modeling and computational methods for multi-scale …qinli/research/qin_thesis.pdfthis dissertation...
TRANSCRIPT
Modeling and Computational Methods for
Multi-scale Quantum Dynamics and Kinetic
Equations
By
Qin Li
A dissertation submitted in partial fulfillment of the
requirements for the degree of
Doctor of Philosophy
(Mathematics)
at the
UNIVERSITY OF WISCONSIN – MADISON
2013
Date of final oral examination: May 6, 2012
The dissertation is approved by the following members of the Final Oral Committee:
Professor Shi Jin, Professor, Mathematics
Professor Alex Kiselev, Professor, Mathematics
Professor Saverio Spagnolie, Assistant Professor, Mathematics
Professor Samuel Stechmann, Assistant Professor, Mathematics
Professor Jean-Luc Thiffeault, Associate Professor, Mathematics
i
Abstract
This dissertation consists of two parts: quantum transitions (Part 1) and hydrodynamic limits
of kinetic equations (Part 2). In both parts, we investigate the inner mathematical connections
between equations for different physics at different scales, and use these connections to design
efficient computational methods for multi-scale problems.
Despite its numerous applications in chemistry and physics, the mathematics of quantum
transition is not well understood. Using the Wigner transformation, we derive semi-classical
models in phase space for two problems: the dynamics of electrons in crystals near band-
crossing points; surface hopping of quantum molecules when the Born-Oppenheimer approx-
imation breaks down. In both cases, particles may jump between states with comparable
energies. Our models can capture the transition rates for such processes. We provide analytic
analysis of and numerical methods for our models, demonstrated by explicit examples.
The second part is to construct numerical methods for kinetic equation that are efficient
in the hydrodynamic regime. Asymptotically, the kinetic equations reduce to fluid dynamics
described by the Euler or Navier-Stokes equations in the fluid regime. Numerically the Boltz-
mann equation is still hard to handle in the hydrodynamic regime due to the stiff collision
term. We review the theoretical work that links the two sets of equations, and present our
asymptotic-preserving numerical solvers for the Boltzmann equation that naturally capture
the asymptotic limits in the hydrodynamic regime. We also extend our methods to the case of
multi-species systems.
ii
Acknowledgements
My deepest thanks go to my advisor, Professor Shi Jin. Without three and a half years of
discussions with him, as well as his advice and support, this dissertation would not exist.
Thank you!
I would also like to thank all my committee members for reading this work and providing
their invaluable comments and suggestions.
I owe a debt of gratitude to my collaborators: Prof. Lorenzo Pareschi, Prof. Xu Yang and
Dr. Lihui Chai, and I look forward to our discussions in the future. I am grateful to Professors
Jose Carrillo, Laurent Desvillettes, Alex Kiselev, Jian-Guo Liu, Paul Milewski, James Ross-
manith and Eitan Tadmor who have been constant sources of help and encouragement for me.
Thanks to Bokai Yan, Jingwei Hu and Li Wang for sharing their insights.
Thanks to Weiwei & Canran, Suky & Tommy, Jo, Keith, Lalit, Lee, Liu, Mike, Sole,
Yongqiang and Zhennan for making these years count. I could never capture in this space
what you each mean to me, but I do appreciate you guys for being a part of my life.
One of the hardest things about these four years has been being 1500 kilometers from
Yang-Le. Waiting for your phone calls has been very painful in most of the days. I was not
able to see you much, but your passion for both science and life has always been with me, and
your love has made my life immeasurably better. Finally, and most importantly, I thank my
parents for making any of this possible. I don’t know how you figured it out but you have been
and will alway be brilliant parents.
iii
List of Figures
2.3.1 The eigenvalues of the Mathieu Model, V (x) = cosx. . . . . . . . . . . . . . . . 38
2.3.2 Periodic potential problem. Linear U(x). t = 0.5. The left and right column
are for the position density ρε, and c.d.f. γε respectively. The solid line, the
dash line and the dotted line are the numerical solutions to the Schrodinger, the
Liouville-A and the Liouville-S respectively. . . . . . . . . . . . . . . . . . . . . 45
2.3.3 Periodic potential problem. Linear U(x). Time evolution of m1(t) defined in
(2.3.10). . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 46
2.3.4 Periodic potential problem. Linear U(x). Errε as function of ε at t = 0.5. . . . 47
2.3.5 Periodic potential problem. Example 1. t = 0.5. The left and right columns
show the position density ρε, and the c.d.f. γε respectively. . . . . . . . . . . . 48
2.3.6 Periodic potential problem. Example 1. Time evolution of m1(t) defined in
(2.3.10). . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 49
2.3.7 Periodic potential problem. Example 1. The L1 error Errε between Liouville-A
and the Schrodinger solution at t = 0.5. . . . . . . . . . . . . . . . . . . . . . . 50
2.3.8 Periodic potential problem. Example 2. At time t = 0.5. The left and the right
columns show the position density ρε, and the c.d.f. γε respectively. . . . . . . 52
2.3.9 Periodic potential problem. Example 2. The L1 error Errε as a function of ε
at t = 0.5. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 53
3.3.1 Surface hopping problem. Example 1. The left column is for ρ+schr/liou, the
density on the upper band, and the right column gives the results of ρ−schr/liou.
∆ =√ε/2. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 80
iv
3.3.2 Surface hopping problem. Example 1. Time evolution of the population (defined
in (3.3.1) and (3.3.2)) on the upper band P+schr/liou. ∆ =
√ε/2. . . . . . . . . . 81
3.3.3 Surface hopping problem. Example 1. Errε decreases at O(ε). t = 10√ε. . . . . 82
3.3.4 Surface hopping problem. Example 2. ∆ =√ε/2. The left column is the density
on the upper band ρ+schr/liou and the right column is ρ−schr/liou. . . . . . . . . . . 83
3.3.5 Surface hopping problem. Example 2. Errε as a function of ε at t = 10√ε. . . . 84
3.3.6 Surface hopping problem. Example 3. Long time behavior of P+. . . . . . . . . 84
3.3.7 Surface hopping problem. Example 4. The left column is for ρ+schr/liou, the
density on the upper band, and the right column gives the results of ρ−schr/liou.
δ =√ε/2. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 85
3.3.8 Surface hopping problem. Example 4. Time evolution of the population on
the upper band P+schr/liou. δ =
√ε/2. The legend “Schrodinger” represents the
solution of the Schrodinger equation, “Liouville-j” represents the solution of the
Liouville system by the domain decomposition method with ∆x = ∆p = h in
the classical regions and ∆x = ∆p = h/2 in then semi-classical region, where
h =√ε/2j−1, and j = 1, 2, 3, 4 . . . . . . . . . . . . . . . . . . . . . . . . . . . 86
3.3.9 Surface hopping problem. Example 4. Errε at t = 10√ε. The Liouville system
by the domain decomposition method with ∆x = ∆p =√ε/8 in the classical
regions and ∆x = ∆p =√ε/16 in then semi-classical region. . . . . . . . . . . . 87
4.1.1 Velocity change after the collision. . . . . . . . . . . . . . . . . . . . . . . . . . 93
v
5.2.1 ExpRK method. Convergence rate test. In each picture, 4 lines are plotted:
the lines with dots, circles, stars and triangles on them are given by results
of ExpRK2-F, ExpRK2-V, ExpRK3-F and ExpRK3-V respectively. The left
column is for Maxwellian initial data, and the right column is for initial data
away from Maxwellian (5.2.37). Each row, from the top to the bottom, shows
results of ε = 1, 0.1, 10−3, 10−6 respectively. . . . . . . . . . . . . . . . . . . . . 125
5.2.2 Exp-RK method. The Sod problem. Left column: ε = 0.01. The solid line
is given by explicit scheme with dense mesh, while dots and circles are given
by ExpRK2-F and ExpRK2-V respectively, both with Nx = 100. h = ∆x/20
satisfies the CFL condition with CFL number being 0.5. Right column: For
ε = 10−6, both methods capture the Euler limit. The solid line is given by the
kinetic scheme for the Euler equation, while the dots and circles are given by
ExpRK2-F and ExpRK2-V. They perform well in rarefaction, contact line and
shock. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 127
5.2.3 Mixing Regime: ε(x) . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 128
5.2.4 ExpRK method. The mixing regime problem. The left column shows compari-
son of RK2 and RK3 using the ExpRK-V. The solid line is the reference solution
with a very fine mesh in time and ∆x = 0.0025, the dash line is given by RK3
and the dotted line is given by RK2, both with Nx = 50 points. The right col-
umn compare two methods, both given by RK3, with the reference. The dash
line is given by ExpRK-V, and the dotted line is given by ExpRK-F. Nx = 50
for both. h is chosen to satisfy CFL condition, in our case, the CFL number is
chosen to be 0.5. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 130
6.2.1 Distribution function for disparate masses system. The distribution function for
heavy species is much narrower than that of the light species. . . . . . . . . . . 149
vi
6.2.2 Multi-species Boltzmann equation. BKG-penalization method. The stationary
shock problem. As ε→ 0, solution of the Boltzmann equation goes to the Euler
limit. t = 0.1. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 155
6.2.3 Multi-species Boltzmann equation. BKG-penalization method. The stationary
shock problem. ε = 0.1, t = 0.1. The dashed line is given by the AP scheme,
and the solid line is given by the forward Euler with a fine mesh: ∆x = 0.01
and h = 0.0005. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 156
6.2.4 The stationary shock problem. ε = 10−5, δf = f1 −M1 diminishes as time
evolves. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 156
6.2.5 Multi-species Boltzmann equation. BKG-penalization method. The stationary
shock problem. u1 and u2 at x = −0.5 on the left and T1, T2 on the right, as
functions of time, for ε = 10−2 and 10−5 respectively. Note the different time
scales for the two figures. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 157
6.2.6 Multi-species Boltzmann equation. BKG-penalization method. The Sod prob-
lem. Indifferentiability. t = 0.1. ρ, u and T are computed using two species
model (“o”) and one species model (“.”). . . . . . . . . . . . . . . . . . . . . . 157
6.2.7 Multi-species Boltzmann equation. BKG-penalization method. The Sod prob-
lem. t = 0.1. As ε→ 0, the numerical results go to the Euler limit. Solid lines
give the Euler limit computed by using [142]. . . . . . . . . . . . . . . . . . . . 158
6.2.8 Multi-species Boltzmann equation. BKG-penalization method. The Sod prob-
lem. t = 0.1. For ε = 0.1, 0.01, we compare the results of the AP scheme, given
by the circled lines, and the results of the forward Euler with a fine mesh, given
by the solid lines. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 159
6.2.9 Multi-species Boltzmann equation. BKG-penalization method. The Sod prob-
lem. ε = 10−4. δf = f1 −M1. . . . . . . . . . . . . . . . . . . . . . . . . . . . . 159
vii
6.2.10Multi-species Boltzmann equation. BKG-penalization method. The Sod prob-
lem. The three figures show velocities u1 and u2 at x = −0.3 as functions of
time for ε = 0.1, 0.01, and 10−5 respectively. Note different time scales for three
figures. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 160
6.2.11Multi-species Boltzmann equation. BKG-penalization method. The disparate
masses problem. The left figure shows the initial distributions and the right
figure show the time evolutions of fH and fL. fH is put at the top and fL is at
the bottom. At t = 0.007, fL is close to ML while fH is still far away from the
equilibrium. Note the different scales for v. . . . . . . . . . . . . . . . . . . . . 160
6.2.12Multi-species Boltzmann equation. BKG-penalization method. The disparate
masses problem. The velocities for the two species converge to each other. . . . 161
6.3.1 Multi-species Boltzmann equation. ExpRK method. The stationary shock prob-
lem. We compare the numerical results at t = 0.1 by the Exp-RK method (the
dotted line), the BGK penalization method (the circled line), and the reference
solution (the solid line). The left figures are for ε = 1, and the right ones are
for ε = 0.1. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 167
6.3.2 Multi-species Boltzmann equation. ExpRK method. The stationary shock prob-
lem. As ε goes to zero, the results get to the Euler limit, i.e. a stationary shock. 168
6.3.3 Multi-species Boltzmann equation. ExpRK method. The stationary shock prob-
lem. The velocities converge to the mean velocity at different rates for different
ε’s. The circles and dots stand for the velocities of the first and second species
respectively. The smaller ε gives the faster convergence rate. . . . . . . . . . . . 168
viii
6.3.4 Multi-species Boltzmann equation. ExpRK method. The Sod problem. Indis-
tinguishability. The solid lines are the results of treating the system as single
species, and they agree with the results given by the dashed lines, which are the
results of a system that has two species with the same mass. . . . . . . . . . . . 169
6.3.5 Multi-species Boltzmann equation. ExpRK method. The Sod problem. At
t = 0.1, we compare the results give by the Exp-RK method (the dotted line),
the BGK penalization method (the circled line) and the reference solution (the
solid line). The left figures are for ε = 1, and the right ones are for ε = 0.1. . . 169
6.3.6 Multi-species Boltzmann equation. ExpRK method. The Sod problem. As ε
goes to zero, the numerical results by the Exp-RK method converge to the Euler
limit. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 169
6.3.7 Multi-species Boltzmann equation. ExpRK method. The Sod problem. The
velocities converge to the mean velocity at different rates for different ε’s. The
circled line stands for species one, and the dotted line stands for species two.
The smaller ε gives the faster convergence rate. . . . . . . . . . . . . . . . . . . 170
6.3.8 Multi-species Boltzmann equation. ExpRK method. The Sod problem. f −M
for ε = 1 and ε = 10−6 at t = 0.1. . . . . . . . . . . . . . . . . . . . . . . . . . . 170
ix
Contents
Abstract i
Acknowledgements ii
I Quantum Transition 1
1 Semi-classical models 3
1.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 3
1.1.1 WKB approximation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 5
1.1.2 Gaussian beam method . . . . . . . . . . . . . . . . . . . . . . . . . . . 7
1.2 Wigner transform . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 8
1.2.1 Properties . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 10
1.2.2 Wigner equation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 11
2 Semi-classical models for crystals 13
2.1 Dynamics of electrons in crystals . . . . . . . . . . . . . . . . . . . . . . . . . . 14
2.1.1 Non-dimensionalization . . . . . . . . . . . . . . . . . . . . . . . . . . . 15
2.1.2 Spectrum analysis . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 16
2.2 Semi-classical model in phase space . . . . . . . . . . . . . . . . . . . . . . . . . 20
2.2.1 Model derivation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 21
2.2.2 Model derivation – symmetric case . . . . . . . . . . . . . . . . . . . . . 30
2.2.3 Main properties of the model . . . . . . . . . . . . . . . . . . . . . . . . 32
2.2.4 Numerical method . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 36
x
2.3 Numerical example . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 37
2.3.1 Linear U(x) . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 38
2.3.2 A domain decomposition computation . . . . . . . . . . . . . . . . . . . 44
3 Semi-classical description of molecular dynamics 54
3.1 Born-Oppenheimer approximation and problem background . . . . . . . . . . . 57
3.1.1 Equation set up . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 58
3.1.2 Basic properties . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 62
3.2 Semi-classical models . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 63
3.2.1 Weyl quantization and Wigner transform . . . . . . . . . . . . . . . . . 64
3.2.2 Adiabatic and non-adiabatic models . . . . . . . . . . . . . . . . . . . . 67
3.2.3 Properties and numerical methods . . . . . . . . . . . . . . . . . . . . . 73
3.3 Numerical examples . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 76
3.3.1 Example 1: 1D, pure state initial data . . . . . . . . . . . . . . . . . . . 77
3.3.2 Example 2: 1D, mixed state initial data . . . . . . . . . . . . . . . . . . 77
3.3.3 Long time behavior . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 78
3.3.4 Example 4: 2D pure initial data . . . . . . . . . . . . . . . . . . . . . . 78
II Kinetic Theory 88
4 Kinetic theory and the hydrodynamic limit 91
4.1 Boltzmann equation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 91
4.1.1 Properties of the collision term . . . . . . . . . . . . . . . . . . . . . . . 94
4.1.2 Non-dimensionalization and rescaling process . . . . . . . . . . . . . . . 96
4.2 Asymptotic limit . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 98
xi
5 Asymptotic preserving numerical methods 101
5.1 BKG-penalization method . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 104
5.2 Exponential Runge-Kutta method . . . . . . . . . . . . . . . . . . . . . . . . . 106
5.2.1 Numerical methods . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 107
5.2.2 Properties of ExpRK schemes . . . . . . . . . . . . . . . . . . . . . . . . 112
5.2.3 Numerical example . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 122
6 Multi-species systems 131
6.1 The multi-species system . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 131
6.1.1 The multi-species Boltzmann equation . . . . . . . . . . . . . . . . . . . 132
6.1.2 Properties of the multi-species Boltzmann equation . . . . . . . . . . . . 133
6.1.3 A BGK model . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 136
6.2 BGK-penalization method . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 138
6.2.1 Numerical method . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 139
6.2.2 The AP property of the time discretization . . . . . . . . . . . . . . . . 143
6.2.3 Disparate masses . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 148
6.2.4 Numerical examples . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 151
6.3 ExpRK method for the multi-species Boltzmann equation . . . . . . . . . . . . 154
6.3.1 Exponential Runge-Kutta method . . . . . . . . . . . . . . . . . . . . . 154
6.3.2 Positivity and asymptotic preserving properties . . . . . . . . . . . . . . 163
6.3.3 Numerical example . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 166
A Quantum system 172
A.1 Some basic analysis of the semi-classical Liouville systems . . . . . . . . . . . . 172
A.1.1 Weak convergence . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 172
A.1.2 Strong convergence: for constant b . . . . . . . . . . . . . . . . . . . . . 173
xii
A.2 The integration method of a simple model system . . . . . . . . . . . . . . . . . 175
A.3 The derivations of (3.2.18) . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 177
B Kinetic theory 180
B.1 Positivity of the mass density in ExpRK-V . . . . . . . . . . . . . . . . . . . . 180
B.2 ‖P (f)− P (g)‖ ≤ ‖f − g‖ for single species . . . . . . . . . . . . . . . . . . . . . 182
B.2.1 In d2 norm . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 182
B.2.2 In L2 norm . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 184
B.3 ‖P (f)‖ < C‖f‖ multi-species . . . . . . . . . . . . . . . . . . . . . . . . . . . . 185
B.3.1 In d2 norm . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 185
B.3.2 In L2 norm . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 187
Bibliography 188
Part I
Quantum Transition
1
2
In this part, we study quantum systems with discrete spectrum, and explore semi-classical
models and numerical methods to compute quantum transitions between energy states. We
start with first principles: given a Schrodinger equation with an external potential, one first
solves the eigenvalue problem of the Hamiltonian operator. The eigenfunctions are usually
referred to as molecular orbitals in quantum chemistry, and the eigenvalues are called energy
bands in physics. The solution to the Schrodinger equation is a superposition of all these
eigenstates, with the coefficients evolving in time. Most of the time, the energy states are
decoupled, which means their coefficients evolve independently. However, in some cases, eigen-
values get very close to each other, or even cross. In physics, when this happens, one could
not distinguish two energy states that share the same energy value. And mathematically, the
projection coefficients need to be modified in response to this quantum phenomena. Quantum
transition is a very important issue because it is on the pathway for most chemical reactions.
Qualitatively how transition functions and gets involved in chemical reactions has been studied
in depth, but not much is known quantitatively due to the lack of mathematical tools avail-
able. We address such issues from the perspective of semi-classical limits in phase space, and
correspondingly we design efficient numerical methods that give accurate results. Quantum
transition occurs in many quantum problems, but we only deal with the following two: the
first is band crossing problem that appears in the dynamics of electrons in crystals and the
second is the surface hopping problem in quantum molecular dynamics.
In the following, we firstly briefly go over the basic ideas of semi-classical limit in chapter 1,
with an outline of the currently dominating methods. We will focus on the Wigner transform,
the one we rely on in the derivation of our model. Chapter 2 and 3 are devoted to the two
problems mentioned above respectively.
3
Chapter 1
Semi-classical models
In this chapter we briefly go over the standard methods developed in the past that give the
semi-classical limit for quantum systems. Techniques on dealing with quantum transitions are
not going to be addressed in this chapter.
1.1 Introduction
The Schrodinger equation is a canonical dispersive equation, whose waves of different wave-
length propagate at different speeds. Consider the Cauchy problem for the Schrodinger equa-
tion:
iεψε = −ε2
2∆ψε + V ψε, (t, x) ∈ R+ × Rd (1.1.1)
ψε(x, t = 0) = ψεin(x)
ψε(t, x) is the wave function, depending on time and space. It can be either a scalar or
vector. The space is d dimensional. ε denotes the semi-classical parameter, usually indicating
the microscopic/macroscopic ratio. It is assigned different physical meanings depending on
different problems in different regimes, as will be clear later on, and for now we only assume
it is a dimensionless small parameter. V is the potential term. Associated with the dimension
of the wave function, it could be either a scalar or a matrix. Most of the time it relies on both
time and space, but in this thesis, its time dependence is not going to be discussed. When
4
V depends on ψ, nonlinear effects also appear, and the most classical model in this case is
V = ±|ψ|2. This situation will be omitted from this thesis as well.
For later use, we define the following:
• Hamiltonian operator:
H = −ε2
2∆ + V (x). (1.1.2)
This Hamiltonian is consistent with the classical one H = p2
2 +V (x), whose p is replaced
by the momentum operator p = −iε∇x.
• Physical observables:
particle density: nε(x, t) = |ψε|2, (1.1.3a)
current density: Jε(x, t) = − iε
2
(ψε∇xψε − ψε∇xψε
)(1.1.3b)
energy density: eε(x, t) =1
2|ε∇xψε|2 + V nε (1.1.3c)
In practice only these physical observables could be measured and are cared about.
Simple derivation shows
∂tnε(x, t) +∇x · Jε(x, t) = 0 (1.1.4)
In the absence of V , its analytical solution is plane wave:
ψε(t, x) =
∫C(ξ) exp
(i
ε
(ξ · x− t
2|ξ|2))
dξ (1.1.5)
where C(ξ) is the projection onto the mode ξ ∈ Rd and they are determined by the initial
data. Apparently the solution features ε oscillations in both time and space. For fixed ε, the
development of numerical methods is very mature, both for time-dependent and independent
case, however, in the semi-classical regime, as ε diminishes, very fine mesh is required according
5
to the typical numerical analysis, and thus leads to extremely heavy computational cost.
To overcome this difficulty, we study the asymptotic behavior of the solution and include the
asymptotic mathematical analysis while designing numerical schemes to save some numerical
cost. Several methods are in order: WKB expansion, Gaussian beam method, the Hagedorn
expansion, and the Wigner transformation. Basically all these methods assume a certain ansatz
for the equation that take the oscillations into account, and one retrieves the leading order
and sometimes the second order term as ε → 0. We briefly review these asymptotic methods
in the following, and one could find more details in [115] and the references therein. Wigner
transform method belongs to this asymptotic analysis category as well, but it will be discussed
in details in the following section.
1.1.1 WKB approximation
WKB method, also known as WKBJ method sometimes, is named after Wentzel-Kramers-
Brillouin (Jeffreys when J is included). It is a general asymptotic method for linear high order
differential equations. In case of time-dependent Schrodinger equation, the ansatz writes as:
ψε = aε(t, x) exp (Sε(t, x)) (1.1.6)
Assume the two real-valued functions: the amplitude aε(x, t) and Sε(x, t) have asymptotic
expansion:
aε = a+ εa1 + ε2a2 + · · · (1.1.7)
Sε = S + εS1 + ε2S2 + · · · (1.1.8)
Plugging this expansion into (1.1.1), by matching orders, one could sequentially solve for the
approximation up to the order desired. To the leading order, S satisfies a Hamiltonian-Jacobi
6
equation with energy given by x2
2 + V (x), while a is governed by a transport equation:
∂tS +1
2|∇xS|2 + V (x) = 0, ∂ta+∇xS∇xa+
a
2∆S = 0. (1.1.9)
The second equation on a could also be interpreted in term of number density ρ = |a|2:
∂tρ+∇ · (ρ∇xS) = 0. (1.1.10)
Given smooth enough S, ρ is governed by a conservation law and could be computed in a
classical way. In this way, the WKB method gives an easy solution to the Schrodinger equation
up to the leading order.
To compute the equation for S is much tricker. In fact, the WKB approximation is only
valid for finite time, since the Hamiltonian-Jacobi equation for S develops singularity even
with smooth initial data: mathematically S could be regarded as integration of some solution
to the Burger’s equation, and it is well-known that even C∞ initial data leads to shocks. By
following characteristics, one solves the following ODE:
x = ξ, ξ = −∇xV. (1.1.11)
and often in time, these characteristics cross, meaning the two particles starting from different
location bump into each other, generating “caustics”.
This phenomenon only occurs for WKB, and not for the original Schrodinger equation,
and thus the “caustics” is artificial and embedded in the definition of WKB ansatz. To fix
the problem beyond the caustics, many techniques can be applied. We mention here the
multiphase WKB approximation, which writes the solution as summation of a series of WKB
form [113, 80], with a phase shift called Keller-Maslov index [124]. For theoretical study, refer
7
to [56]. Level set method is a general way in computing multi-values appeared in Hamiltonian-
Jacobian equation [161], and is adopted here as well [116, 114].
1.1.2 Gaussian beam method
The Gaussian beam method, developed for high frequency linear waves, is an efficient
approximate method that allows accurate computation of the wave amplitude around caus-
tics [168]. It is widely used in geophysics community as well [102]. The validity of its construc-
tion has been mathematically studied by Ralston in [171]. Its usage in quantum system can
be found in [121, 120, 107, 200].
The idea is the following. Assume at the initial time step, the data is given in the form of:
ψε0(t, x, y) = A(t, y)eiT (t,x,y)/ε (1.1.12)
where both the amplitude and the phase function have a new variable y indicating the center
of the beam, as will be more clear soon. T is complex valued function and is Taylor expanded
around y as:
T (t, x, y) = S(t, y) + p(t, y) · (x− y) +1
2(x− y)TM(t, y)(x− y) +O(|x− y|3) (1.1.13)
To form a Gaussian profile, the imaginary part of the Hessian matrix M is positive definite.
In the leading order expansion in both ε and y − x, one gets the ODE system:
dy
dt= p,
dp
dt= −∇yV
dM
dt= −M2 −∇2
yV (1.1.14)
dS
dt=
1
2|p|2 − V, dA
dt= −1
2(Tr(M))A (1.1.15)
8
The equation for M is called Riccati equation, and by following the y-trajectory in this La-
grangian type of approach, one could check several important properties:
1. M is always symmetric and its imaginary part remains being positive definite;
2. The Hamiltonian H = |p|22 + V (x) is conserved;
3. A(t, y) does not blow up.
The first property guarantees that the profile is always a Gaussian function, while the second
and the third incorporate the physical background.
Given an arbitrary initial data, even if it does not have the form of (1.1.12), it can always
be approximated by the summation of several Gaussian packets, but the number could be as
big as ε−1/2. In case of the WKB initial data, the approximation could be made in a particular
way to reach ε1/2 accuracy [181, 182]. Given these functions, one needs to follow the trajectory
for each packets and evolve them according to (1.1.14). They are summed up at the ending
time. Also see high order extension in [122], and frozen Gaussian beam method in [149, 150].
1.2 Wigner transform
In this section we mainly focus on the Wigner transform developed by Wigner in [193].
Unlike most other asymptotic method, Wigner transform studies the behavior of the solution
in phase space, and helps to establish the link between the quantum wavefunction and the
probability distribution function in classical mechanics. The definition has the uncertainty
principle built in, and thus provides us a powerful tool to untangle the coherence between
the momentum and position in phase space [155]. In representation theory, it is also linked
to Weyl quantization. The latter was developed in 1927 in [192] and is a general framework
that transforms the real phase space functions to its associated Hermitian operators. More
discussion will be found in section 3.2.1.
9
Wigner transform is defined as:
W ε(t, x, k) =
∫Rd
dy
(2π)deikyψε
(t, x− ε
2y)ψε(t, x+
ε
2y). (1.2.1)
where ψε is the complex conjugate of ψε, with ψε solves the Schrodinger equation (1.1.1). The
multiplication of two wave functions ψε(x− εy2 )ψε(x+ εy
2 ) is called the density matrix, and the
transform is simply the inverse Fourier transform of it. Note that ψ and its complex conjugate
are evaluated at different position with deviation O(ε), intrinsically reflecting the uncertainty
principle. In bra-ket language, it could be expressed as:
W ε(t, x, k) =
∫Rd
dy
(2π)d〈x+
ε
2y|ρ|x− ε
2y〉eiky. (1.2.2)
The definition is to characterize the probability of finding a particle at time t, location x with
momentum k, as a resemblance to the probability distribution function in statistical mechanics.
Remark 1.2.1. Note that both two formulas above are written in x-presentation, consistent
with the analysis and numerical computation that will be carried below. p-presentation could
be used as well. One only needs to replace the states |x± ε2y〉 by |p± ε
2q〉 and integrate over q,
the deviation in the momentum space. The discussion for that is neglected from here.
As a resemblance to the distribution function on the phase space in statistical mechanics,
this Wigner function is introduced to characterize the probability of finding a particle at time
t, location x with momentum k. However, a classical particle has a definite position and
momentum, and hence it is represented by a point in phase space, in the quantum case, due
to the uncertainty principle, this restriction breaks down, and the Wigner function can only
be understood in a quasi-probability manner, and thus present some properties that are not
shared by conventional probability distribution.
10
1.2.1 Properties
Several properties are in order.
1. W ε is a real-valued function.
Take the conjugate of the definition (1.2.1) it is easily seen that W ε = W ε.
2. Time symmetry.
ψε → ψε ⇒W (t, x, k)→W (t, x,−k). (1.2.3)
It is called time symmetry simply because k indicates the momentum and W ε(t, x,−k)
is the wave W ε(t, x, k) travelling backward.
3. ψε cannot be fully recovered.
By definition, W ε has a quadratic form, and if one shift the phase of ψε by a constant θ,
i.e. ψε → ψεeiθ, W ε is unchanged.
4. The physical observables can be recovered from the moments:
particle density: nε(x, t) =
∫RdW εdk (1.2.4a)
current density: Jε(x, t) =
∫RdkW εdk (1.2.4b)
energy density: eε =
∫Rd
( |k|22
+ V (x)
)W εdk (1.2.4c)
5. Wigner function does not preserve positivity.
(a) This property distinguishes the Wigner function from the conventional density dis-
tribution function. The reason behind it is the uncertainty principle. A classical
particle has a definite position and momentum, and hence can be represented by a
certain point on the phase space, but in the quantum case, a particle could only be
11
specified in space and momentum simultaneously with an error of O(~) (ε in this
thesis).
(b) If the Wigner function is smoothed through a filter with size bigger than ε, positivity
could still be obtained. The most classical way is Husimi transform [129] where a
Gaussian function with width of O(ε) is used to convolute with the density matrix.
1.2.2 Wigner equation
By plugging the Schrodinger equation into the definition of Wigner function, one has the
Wigner equation:
∂tWε + k · ∇xW ε =
1
iε
∫dω
(2π)deiωxV (ω)
[W ε(x, k − εω
2)−W ε(x, k +
εω
2)]. (1.2.5)
Here V is the Fourier transform of the potential V . Assume V (x) is a smooth enough function
with ∇xV = O(1), and by performing asymptotic expansion for W ε = W0+εW1+· · · , formally
one obtains the classical Liouville equation:
∂tW0 + k · ∇xW0 −∇xV · ∇kW0 = 0 (1.2.6)
This has been proved in [151, 147]:
Theorem 1.2.1. Suppose ψε is uniformly bounded in L2(Rd) with respect to ε:
sup0<ε≤1 |ψε|L2 <∞, ∀t, (1.2.7)
then the associated Wigner function W ε ∈ S′(Rd × Rd) is weak-* compact. There exists one
12
sequence such that:
W ε ε→0−−−→W0 in L∞(
[0, T ];S′(Rd × Rd
))weak-* (1.2.8)
The limit W0 is called the Wigner measure. It is positive and follows the Liouville equa-
tion (1.2.6).
Remark: One way to prove the positivity of this limit is to link it with the limit of Husimi
measure, which keeps being positive for all ε [151].
Follow the characteristics of the equation (1.2.6), the resemblance to the classical mechanics
is easily seen:
x = k, k = −∇xV (x). (1.2.9)
This tells us that k is the velocity of a particle, and k, the acceleration is governed by the
derivative of the potential V (x), which is exactly the Newton’s 2nd Law. In this sense, this
two different types of physics are linked with an error of O(ε). Compared with the WKB
approximation, this approach unfold the singularities (caustics) and does not give the exact
wave function ψε, and the phase is globally well-defined.
Wigner transform is the tool we use to investigate the two problems described in the
following two chapter.
13
Chapter 2
Semi-classical models for crystals
In this chapter we investigate the dynamics of electrons in crystals in the semi-classical
regime. We explore the mathematical description for the quantum transition between energy
bands, and develop efficient semi-classical numerical solver.
This problem is difficult because in the regime we consider, the lattice potential oscillates
at the same scale of the rescaled Planck constant, making the numerical computation very
expensive. Using the Bloch-Floquet theory, one could obtain the spectrum of the Hamiltonian
by solving the associated eigenvalue problem. When the eigenvalues are very well separated
from each other, it is proved that in the semi-classical limit, in phase space, the projection
onto different bands are decoupled, and each coefficient evolves independently according to a
simple transport equation. However, when the eigenvalues are degenerate, little theory on the
quantum transition is known.
We solve this problem by performing the Wigner transform on the Schrodinger equation
and study the resulting Wigner equation in phase space. To better interpret the structure of
the phase space, we also perform the Wigner transform onto the Bloch eigenfunctions. This
gives us the orthogonal basis in phase space. We project the Wigner function onto them, and
discovered that the off-diagonal entries of the obtained Wigner matrix naturally reflect the
quantum transition rates. The associated set of equations are obtained. This new model is
asymptotically consistent with the classical model under the assumption that the energy bands
are non-degenerate. Numerically, the story has two sides: on the one hand, the new model gives
14
full information both around and far away from the band-crossing region, but its oscillatory
nature generates expensive numerical cost, and on the other hand, the classical model, though
could not provide transition rate around the crossing point, gives sufficient information away
from it. With this observation, we construct a domain-decomposition numerical method: we
decompose the phase space into separate domains, and compute proper equations in different
regions and hybridize them by appropriate boundary conditions. In this way we capture the
correct transition rate without sacrificing much numerical expense.
In section 2.1, we briefly go over the setup of the problem and introduce its basic properties.
The standard technique to tackle the problem is the Bloch-Floquet theory. In section 2.2 we go
over the classical Wigner transform approach, and around the energy band crossing point where
the classical model breaks down, we develop a new model that captures the quantum transition.
Its basic properties will be presented in subsection 2.2.3 and the associated numerical methods
are designed in subsection 2.2.4. For simplicity both the analysis and the computation are done
in one dimension but they could be easily extended to higher dimensions. We show several
numerical examples in section 2.3.
2.1 Dynamics of electrons in crystals
A crystal is a solid material whose atoms are arranged in an ordered pattern extending
in spatial domain. In our problem we only care about the dynamics of electrons, and all
bigger molecules are considered at rest, leaving electrons a static potential. Based on this set
up, the Schrodinger equation used contains two potential terms: the lattice periodic potential
generated by the ionic cores, and the experimentally imposed external potential that is usually
introduced by electromagnetic fields. The latter typically vary on much larger spatial scales.
15
Denote l as the lattice constant, we rewrite (1.1.1) as:
i~∂tψ = −~2
2∆ψ + V
(xl
)ψ + U(x)ψ. (2.1.1)
Here the V (x) part in the potential term reflects the lattice oscillation and U(x) is the smooth
external potential. ~ is the Planck constant.
Two small parameters are involved in the equation: the Planck constant ~ and the lattice
constant l. Depending on the regimes one is interested in, the parameters need to be rescaled
and balanced, resulting different behavior of the solution. We firstly perform the nondimen-
sionalization in section 2.1.1, and pick a regime where resonance of lattice potential could be
seen. Basic properties, especially the Bloch-Floquet theory will be introduced in 2.1.2.
2.1.1 Non-dimensionalization
Define τ = l2/~ as the quantum time scale, and denote T and L as macroscopic time and
length scale. Rescale the parameters [10, 169, 57]:
x =x
L, t =
t
T, ε =
l
L, h =
~TL2, (2.1.2)
and the two potential functions:
V (x
l) =
τ2
ml2V (x), U(
x
L) =
T 2
L2U(x). (2.1.3)
After dropping the tildes, one gets:
ih∂tψ = −h2
2∆ψ +
h2
ε2V(xε
)ψ + U(x)ψ. (2.1.4)
16
This equation is nondimensionalized, and it shows the competition between two parameters h
and ε. As one can see, if ε is big compared to h, one is in a regime that no periodic structure of
the crystal cannot be seen, and the system is roughly classical; if ε is comparably much smaller,
the homogenization limit ε→ 0 of the periodic structure is performed faster than the classical
limit h → 0. In fact it is proved that this limit yields a semi-classical Vlasov equation with
parabolic bands defined by some effective mass tensor [169] when energy bands are assumed
to be separated. The most interesting case is when ε and h is at the same order. If so, the
periodic potential V (x) and the external potential U(x) are of the same order, and one could
see the O(1) impact of the periodic structure to the system and expect some resonance. In
this case, write h = ε and the equation becomes:
iε∂tψε = −ε
2
2∆ψε + V
(xε
)ψ + U(x)ψε, (2.1.5)
ψε(t = 0, x) = ψεI(x). (2.1.6)
For simplicity, we assume V (x) is 2π periodic, and U(x) is smooth and of O(1).
Asymptotic analysis has been extensively studied in, e.g., [162, 89, 164, 10], relying on
different analytical tools, and numerical methods are also developed [79, 81, 82, 105, 106]. We
also mention some papers that study related models is: Free crystal electrons, i.e. U(x) = 0
is studied in [151, 73]. Homogenization limit is studied in [169] with its extension to nonlinear
Coulomb potential in [9]. Stratified potential with periodicity appearing in only one direction
is studies in [11].
2.1.2 Spectrum analysis
It is the periodic potential that oscillates at the scale of ε that generates a great amount of
difficulty in analyzing the equation. To fully make use of the periodicity, the classical way is
to use Bloch-Floquet theory. The theory dates back to Floquet who discovered the underlying
17
mathematics in 1883 [69], and later on Bloch realized its application in the lattice potential
case [17]. The whole theory was summarized in [194]. The idea is to combine the V (xε ) with
the Laplacian, and write it as a new Hamiltonian:
H0 = −ε2
2∆x + V (
x
ε) = −1
2∆z + V (z). (2.1.7)
Here we define z = xε as the fast variable and study the eigenvalue problem:
−1
2∆zΨ(z, p) + V (z)Ψ(z, p) = E(p)Ψ(z, p), (2.1.8a)
Ψ(z + ν, p) = eipνΨ(z, p), ∀ ν ∈ L = 2πZ, (2.1.8b)
∂Ψ
∂z(z + ν, p) = eipν ∂Ψ
∂z(z, p), ∀ ν ∈ L. (2.1.8c)
The boundary conditions for this eigenvalue problem (2.1.8b)-(2.1.8c) is called Bloch boundary
condition. Given a specific p, the solution is equipped with pseudo-periodic boundary condi-
tion, with the absolute value of Ψm unchanged but the phase is shifted by 2πp when z changes
by 2π. With each fixed p, one solves (2.1.8) for a list of discrete eigenvalues Em(p) and the
corresponding Bloch eigenfunctions, marked as Ψm(z, p). Here we assume the multiplicity is
1, i.e. each eigenvalue corresponds to one eigenfunction. Note that Ψm is a function of z with
p regarded either as a parameter or a variable.
We list a few properties as preparation for later analysis[160, 194, 10]:
(a) The eigenvalues Em(p) are 1-periodic in p. Denote L∗ = Z as the lattice for p. The
analysis will be carried on the first Brillouin zone p ∈(−1
2 ,12
). Order the eigenvalues
as E1(p) ≤ E2(p) ≤ ... ≤ Em(p) ≤ .... It can be proved that Em(p) → ∞ as m → ∞,
uniformly in p, and that strict inequality can be achieved outside a zero measure set [194].
(b) Pick one period for p, called the first Brillouin zone B =(−1
2 ,12
), Ψm(·, p) forms a
18
complete orthonormal basis in L2(0, 2π), i.e.
(Ψm,Ψn) :=
∫ 2π
0
dz
2πΨm(z, p)Ψn(z, p) = δmn. (2.1.9)
(c) For all φ ∈ L2(R), one has the following Bloch decomposition:
φ(x) =
∞∑m
∫Bcm(p)Ψm(x, p) dp (2.1.10)
where cm is the Bloch coefficient: cm(p) =∫R φ(x)Ψm(x, p) dx.
Note: Another way to study this problem is to investigate the following function:
Φαm(z, p) = e−ipz Ψα
m(z, p). (2.1.11)
This definition gives a function that is 2π-periodic in z and 1-periodic in p. Φm(·, p) also form
a complete basis in L2(R) space.
As one can see in the property (c), any L2 function can be expanded upon the eigenfunctions
Ψm(z, p). In the absence of the external potential U(x), one could proceed as the following:
• 1st step: Use the expansion (2.1.10) and define the projection onto the mth band:
cm(t, p) =∫ψ(t, z)Ψm(z, p)dz;
• 2nd step: Multiply Ψm(z, p) on both sides of the equation (2.1.5) and integrate over z,
one gets the evolution of cm(t, p):
iε∂tcm =
∫H0ψΨmdz = Em(p)cm. (2.1.12)
The solution could be explicitly written down: cm(t, p) = cm(0, p)eiEmtε .
• 3rd step: Collect all the cm at the ending time.
19
Generally speaking one projects the initial data onto different bands. Since the equation is
linear, the projection coefficients evolve independently in time. cm could be explicit computed,
and thus no heavy computation is needed. However, the method above is only true for zero
external potential U(x). With a smooth nonzero potential U(x), the analysis could be much
more tedious. In fact, in space, we have two scales: x and z, and thus the solution ψ also
contains the slow modes in x and the fast fluctuation in z. So we manually split the two scales:
• 1st step: Separate the scaling and write the solution as ψ = ψ(t, x, z). The Schrodinger
equation thus is revised:
iε∂tψ =
(−∆z
2+ V (z)
)ψε + U(x)ψε −
(ε2∆x
2+ ε∇x · ∇z
)ψε. (2.1.13)
Note that here we change ∇x into ∇x + 1ε∇z.
• 2nd step: Use the expansion (2.1.10) and define the projection onto the mth band:
cm(t, p, x) =∫ψ(t, x, z)Ψm(z, p)dz;
• 3rd step: Multiply Ψm(z, p) on both sides of the equation (2.1.5) and integrate over z,
one gets the evolution of cm(t, p, x):
iε∂tcm = −ε2∆x
2cm + (Em(p) + U(x)) cm + ε
∑n
∫∇xcn
∫Rd
Ψn∂zΨmdzdp. (2.1.14)
The evolution in time has two parts on the right: the first part is given by the first two
terms. They resemble the Schrodinger Hamiltonian, containing both the kinetic energy
− ε2∆x2 and the potential Em(p) + U(x). One thing to be noted here is that different
energy bands provide different potential terms – i.e. different strength of acceleration.
The second part is the third term on the right: it is of O(ε), and has the Berry connection
term involved:∫
Ψn∇zΨmdz [12]. It shows the connection between different bands.
20
• 4th step: Solve (2.1.14) for cm and collect all the modes at the ending time.
Apparently (2.1.14) is no easier than the original Schrodinger equation: it is not only oscillatory
in both time and space, but also shows strong nonlinear coupling among bands.
All in all, when U(x) is present, the solution has two scalings and the standard Bloch-
Floquet theory does not simplify the computation. Nevertheless this theory provides us one
way to decompose the L2 function space, and will be used in our Wigner approach.
2.2 Semi-classical model in phase space
In this section we derive a model for the crystal problem in the framework of Wigner
transformation. In fact if one follows the standard Wigner transform steps, and define the
Wigner equation as in (1.2.1), it is easy to obtain the following Wigner equation:
∂tWε + k∂xW
ε =1
iε
∑µ∈L∗
eiµx/εV (µ)[W ε(x, k − µ
2)−W ε(x, k +
µ
2)]
(2.2.1)
+1
iε
∫R
dω
2πeiωxU(ω)
[W ε(x, k − εω
2)−W ε(x, k +
εω
2)],
where U(ω) and V (µ) are defined as:
U(ω) =
∫R
dy e−iωyU(y), V (µ) =1
2π
∫ 2π
0dy e−iµyV (y). (2.2.2)
with ω ∈ R, and µ ∈ L∗. Note V (x) is periodic and thus it only has Fourier series instead of
Fourier transform.
On the right hand side, the second term is approximated by ∂xU∂kW . The first term,
however, can be much trickier. This term is a series summation instead of an integral; the
deviation µ in k is of O(1); and the whole term is multiplied with 1iε , making it impossible for
simplification with standard techniques.
21
2.2.1 Model derivation
To overcome this difficulty, we use the following two new ideas: 1. we separate the fast
variable from the slow one in x, and write function W (t, x, p) as W (t, x, z, p); 2. we found a
basis for W to be expand upon in phase space. The model is carried out with asymmetric
Wigner transform, and the symmetric version will be briefly summarized in the next section.
Define the Wigner function W :
Wε(t, x, k) =
∫R
dy
2πeikyψε (t, x− εy) ψε(t, x). (2.2.3)
It is easy to obtain the Wigner equation (with the fast variable z = xε ):
∂
∂tWε + k
(∂
∂x+
1
ε
∂
∂z
)Wε +
iε
2
(∂
∂x+
1
ε
∂
∂z
)2
Wε
=1
iε
∑µ∈L∗
eiµx/εV (µ) [Wε(x, k − µ)−Wε(x, k)]
+1
iε
∫R
dω
2πeiωxU(ω) [Wε(x, k − εω)−Wε(x, k)] .
(2.2.4)
Under the assumption that U(x) is smooth, high order terms are thrown away, leaving:
∂tWε + k∂xWε − ∂xU∂kWε + i∂2xzWε = −1
εLWε, (2.2.5)
where the skew symmetric operator L is given by
Lf(z, k) = k∂f
∂z+
i
2
∂2f
∂z2− 1
i
∑µ∈L∗
eiµx/εV (µ) [f(x, k − µ)− f(x, k)] .
We use the classical scaling separation argument, asymptotically expand W ε as:
W ε(t, x, k) = W0
(t, x,
x
ε, k)
+ εW1
(t, x,
x
ε, k)
+ · · · (2.2.6)
22
Plug this ansatz into (2.2.4). By balancing both the O(1ε ) and O(1) term, one gets:
LW0 = 0 , (2.2.7a)
∂W0
∂t+ k
∂W0
∂x− ∂U
∂x
∂W0
∂k+ i
∂2W0
∂x∂z= −LW1 . (2.2.7b)
The first equation indicates that W0 is in the kernel of L. To specify the kernel space, and in
particular to seek for a good basis of kerL, Bloch-Floquet theory is adopted [3].
Given two Bloch eigenfunctions Ψm and Ψn, we define the following:
Qmn(z, k) = Qmn(z, µk, pk) =
∫ 2π
0
dy
2πeikyΨm(z − y, pk)Ψn(z, pk), (2.2.8)
where k is an arbitrary real number and is decomposed as:
k = pk + µk, pk ∈ B, µk ∈ L∗. (2.2.9)
Note that this definition is very similar to Wigner transformation (1.2.1), but notice the two
difference: 1. y is integrated over one period 2π; 2. in the wave function z and y have the
same scaling. With this definition, one could prove:
1. Qmn form a basis for W to be expanded upon:
W (t, x, p) =∑mn
σmn(t, x, p)Qmn(z, p). (2.2.10)
It is a direct corollary of the following lemma:
Lemma 2.2.1. Define the inner product 〈·, ·〉:
〈f, g〉 :=∑µ∈L∗
∫ 2π
0
dz
2πf(z, µ)g(z, µ), f, g ∈ L2
((0, 2π), `2(L∗)
),
23
then for any p ∈ (−1/2, 1/2), Qmn(·, ·, p) forms a complete orthonormal basis in
L2((0, 2π)× `2(L∗)
).
Proof. The orthonormal condition
〈Qmn, Qjl〉 = δmj δnl, (2.2.11)
can be proved by simply using (2.1.9). To prove the completeness, it is sufficient to show
that: if there exists an f ∈ L2((0, 2π)× `2(L∗)
), such that 〈f,Qmn〉 = 0 for all m,n ∈ N,
then f(z, µ) ≡ 0. Assume 〈f,Qmn〉 = 0 for all m,n ∈ N. By the definition of Qmn,
∑µ
∫ 2π
0
∫ 2π
0
dydz
(2π)2f(z, µ)eiµyΦm(z − y, p)Φn(z, p) = 0, ∀m,n ∈ N.
Since Φn(·, p) forms a complete orthonormal basis in L2(0, 2π), the above equality
implies that ∑µ
∫ 2π
0
dy
2πf(z, µ)eiµyΦm(z − y, p) = 0, ∀m ∈ N,
thus ∑µ
f(z, µ)eiµy ≡ 0,
which implies that f(z, µ) ≡ 0.
2. A straightforward computation gives:
LQmn(z, k) = LQmn(z, µ, p) = i(Em(p)− En(p))Qmn(z, µ, p) . (2.2.12)
Apparently the kernel of L highly relies on the structure of energy bands. Before carrying out
any further derivations, we make the following assumption.
24
Assumption 2.2.1. Eigenvalues are strictly apart from each other everywhere in p, namely
E1(p) < E2(p) < ... < Ej(p) < ....
This assumption may be valid or not, and depending on its validity, we call the scenario
adiabatic case or non-adiabatic case, and derive two different models respectively, as shown
below. Associated with it, we also compute the initial data.
Adiabatic case
We derive the model for adiabatic case in this part. With the assumption, (2.2.12) clearly
indicates:
kerL = spanQmm,m = 1, 2, · · · . (2.2.13)
According to the leading order equation (2.2.7a), it is easy to see that W0 should be in the
kernel of L, and thus is simply an expansion of Qmm:
W0(t, x, z, k) =∑m
σmm(t, x, p)Qmm(z, µ, p). (2.2.14)
Here σmm representing the expansion coefficients. To close it, we plug the expansion (2.2.14)
back into the second order equation (2.2.7b)), and take the inner product with Qmm on both
sides. The right hand side vanishes due to the skew symmetry of L and (2.2.12):
〈−LW1, Qmm〉 = 〈−W1,LQmm〉 = 0. (2.2.15)
25
The left hand side, on the other hand, gives:
〈∂tW0 + k∂xW0 − ∂xU∂kW0 + i∂xzW0, Qmm〉
=∑n
∂tσnn 〈Qnn, Qmm〉+∑nn
∂xσnn 〈kQnn, Qmm〉
−∂xU(∑
n
∂pσnn 〈Qnn, Qmm〉+∑nn
σm 〈∂pQnn, Qmm〉)
= ∂tσmm − i (∂zΨn, Ψm) ∂xσmm − ∂xU∂pσmm − ∂xU∑n
((∂pΦn, Φm) + (Φn, ∂pΦm)
)= ∂tσmm + ∂pEm∂xσmm − ∂xU∂pσmm . (2.2.16)
By combining (2.2.15) and (2.2.16), one gets transport equation on each band:
∂tσmm + ∂pEm∂xσmm − ∂xU∂pσmm = 0. (2.2.17)
In the derivation of (2.2.16), the following equalities were used:
〈kQmn, Qjl〉 = − i
2
(δnl (∂zΨm,Ψj) + δmj (∂zΨl,Ψn)
), (2.2.18a)
〈∂pQmn, Qjl〉 = δnl (∂pΦm,Φj) + δmj (Φl, ∂pΦn) , (2.2.18b)
〈∂zQmn, Qjl〉 = δnl (∂zΨm,Ψj) − δmj (∂zΨl,Ψn) , (2.2.18c)
∂pEmδmj + (∂pΦm , Φj) (Em − Ej) = −i (∂zΨm , Ψj) , (2.2.18d)
(∂pΦm,Φm) + (Φm, ∂pΦm) = 0. (2.2.18e)
These are simple derivations and the details are omitted.
(2.2.17) is a very important resemblance to the classical Liouville equation (1.2.6) as it also
shows clear connection towards the classical mechanics. The Hamiltonian is H = H0 + U(x),
and on each band, the effective potential is Em(p) + U(x). Following the trajectory, one gets
x = ∂pH = ∂pEm(p) and k = −∂xH = −∂xU(x). The configuration of the energy bands
26
apparently controls the momentum.
Remark 2.2.1. Results for the case that has no external potential were rigorously proved
in [151, 74] in the framework of Wigner series, instead of Wigner transform.
One could also write W =∑
mn σmnQmn, but Qmn with m 6= n only appears in W1,
meaning that σmn ∼ O(ε), negligible in the semi-classical regime.
Non-adiabatic case
In this part we derive the model for the system when Assumption 2.2.1 is turned off.
In fact, when two adjacent energy get very close Em − Em−1 < O(ε), or even ∼ O(ε), the
argument above is no longer valid. With LQm,m−1 = O(ε) in (2.2.6), the previous scaling
separation (2.2.7) is no longer correct, and σm,m−1 should not be eliminated in the leading
order expansion. When σmn with m 6= n acquires O(1) value, we say quantum tunnelling is
observed.
To get the evaluation of these quantities, ones use the full expansion (2.2.10). We plug (2.2.10)
into (2.2.5) and take the inner product with Qmn. The derivation is tedious, and here we only
list the results.
Without loss of generality, we tackle a two-band problem. Define pc = arg minp|E1 − E2|
and assume |E1(pc)− E2(pc)| = O(ε) at pc = 0.
∂σ11
∂t+∂E1
∂p
∂σ11
∂x+
1
i
(∂Ψ1
∂z, Ψ2
)∂σ12
∂x− ∂U
∂x
∂σ11
∂p
=∂U
∂x
[(∂Φ2
∂p, Φ1
)σ21 +
(Φ1 ,
∂Φ2
∂p
)σ12
],
∂σ22
∂t+∂E2
∂p
∂σ22
∂x+
1
i
(∂Ψ2
∂z, Ψ1
)∂σ21
∂x− ∂U
∂x
∂σ22
∂p
=∂U
∂x
[(∂Φ1
∂p, Φ2
)σ12 +
(Φ2 ,
∂Φ1
∂p
)σ21
],
27
∂σ12
∂t+∂E2
∂p
∂σ12
∂x+
1
i
(∂Ψ2
∂z, Ψ1
)∂σ11
∂x− ∂U
∂x
∂σ12
∂p+ i
E1 − E2
εσ12
=∂U
∂x
[(Φ2 ,
∂Φ1
∂p
)σ11 +
(∂Φ2
∂p, Φ1
)σ22 +
(Φ2 ,
∂Φ2
∂p
)σ12 +
(∂Φ1
∂p, Φ1
)σ12
],
∂σ21
∂t+∂E1
∂p
∂σ21
∂x+
1
i
(∂Ψ1
∂z, Ψ2
)∂σ22
∂x− ∂U
∂x
∂σ21
∂p+ i
E2 − E1
εσ21
=∂U
∂x
[(Φ1 ,
∂Φ2
∂p
)σ22 +
(∂Φ1
∂p, Φ2
)σ11 +
(Φ1 ,
∂Φ1
∂p
)σ21 +
(∂Φ2
∂p, Φ2
)σ21
].
This system can be written in vector form as:
∂tσ +A∂xσ +B∂pσ = −BCσ +iD
εσ (2.2.19a)
where
σ = ( σ11 σ12 σ21 σ22 )T (2.2.19b)
B = −∂xU I , D = diag ( 0, E2 − E1, E1 − E2, 0 ) (2.2.19c)
A =
∂pE112ψ12
12ψ21 0
12ψ21 ∂p
E1+E22 0 1
2ψ21
12ψ12 0 ∂p
E1+E22
12ψ12
0 12ψ12
12ψ21 ∂pE2
, (2.2.19d)
C =
0 −φ12 φ21 0
−φ21 φ11 − φ22 0 φ21
φ12 0 φ22 − φ11 −φ12
0 φ12 −φ21 0
, (2.2.19e)
28
ψmn = −i (∂zΨm,Ψn) , and φmn = (∂pΦm,Φn) , (2.2.19f)
Superscript T stands for matrix transpose and I is the identity matrix. Generally, ψ12 = ψ21,
φ12 = −φ21 are complex-valued quantities, and φ11, φ22 are purely imaginary, so A = A† is
Hermitian, and C = −C† is anti-Hermitian. The superscript † denotes the matrix conjugate
transpose.
Initial data for the semi-classical model
One also needs to equip it with appropriate initial condition. Choose the initial data of
the Schrodinger equation as two wave packets along the two Bloch bands in the following form
[31, 105]:
φI = a1(x) Φ1
(xε, ∂xS0(x)
)eiS0(x)/ε + a2(x) Φ2
(xε, ∂xS0(x)
)eiS0(x)/ε. (2.2.20)
Then, the initial data of the Wigner function, for ε 1, has the approximation:
WI(x, z, k) ∼ |a1(x)|2W11(z, k) + |a2(x)|2W22(z, k)
+a1(x)a2(x)(W12(z, k) +W21(z, k)
), (2.2.21)
with
Wmn(z, k) =
∫R
dy
2πeikyΦm(z − y, ∂xS0(x− εy))Φn(z, ∂xS0(x))ei(S0(x−εy)−S0(x))/ε.
Using Taylor expansion on S0(x− εy)− S0(x) and Φm(z − y, ∂xS0(x− εy)), one gets
Wmn(z, k) =
∫R
dy
2πei(k−∂xS0(x))yΦm(z − y, ∂xS0(x))Φn(z, ∂xS0(x)) +O(ε),
29
then by ignoring the O(ε) term, and using the periodicity of Φm(z, p) on z, one can change
the integral into a summation of integrals from 0 to 2π:
Wmn(z, k) =∑ν∈L
∫ 2π
0
dy
2πei(k−∂xS0(x))(y+ν)Φm(z − y, ∂xS0(x))Φn(z, ∂xS0(x)).
Applying the equality ∑ν∈L
eikν =∑µ∈L∗
δ(k + µ),
one gets
Wmn(z, k) =∑µ∈L∗
δ(k + µ− ∂xS0)
∫ 2π
0
dy
2πei(k−∂xS0)yΦm(z − y, ∂xS0)Φn(z, ∂xS0)
=∑µ∈L∗
δ(pk + µ− ∂xS0)
∫ 2π
0
dy
2πeikyΨm(z − y, ∂xS0)Ψn(z, ∂xS0)
=∑µ∈L∗
δ(pk + µ− ∂xS0)
∫ 2π
0
dy
2πeikyΨm(z − y, pk + µ)Ψn(z, pk + µ)
=∑µ∈L∗
δ(pk + µ− ∂xS0)
∫ 2π
0
dy
2πeikyΨm(z − y, pk)Ψn(z, pk).
=∑µ∈L∗
δ(pk + µ− ∂xS0
)Qmn(z, µk, pk). (2.2.22)
In the derivation above, from the second line to the third line, we use the fact that
∫δ(p− p0)f(p0)g(p) dp = f(p0)g(p0) =
∫δ(p− p0)f(p)g(p) dp
to replace the argument ∂xS0 by pk + µ.
Without loss of generality, we assume that ∂xS0 ∈ (−1/2, 1/2), then (2.2.22) becomes
Wmn(z, k) = δ(pk − ∂xS0
)Qmn(z, µk, pk). (2.2.23)
30
Compare (2.2.23) with (2.2.21), one has the initial data for σ:
σ(0, x, p) = δ (p− ∂xS0(x))
(a2
1 a1a2 a1a2 a22
)T. (2.2.24)
2.2.2 Model derivation – symmetric case
This section follows the routine in section 2.2, with the standard symmetric Wigner trans-
form:
W sε (t, x, k) =
∫R
dy
2πeikyφε
(t, x− εy
2
)φε(t, x+
εy
2
).
The derivation is similar, thus we skip the details, and give a list of the results:
1. The Wigner equation corresponding (2.2.4) is
∂W sε
∂t+ k
∂W sε
∂x=
1
iε
∑µ∈L∗
eiµx/εV (µ)[W sε
(x, k − µ
2
)−W s
ε
(x, k +
µ
2
)]+
1
iε
∫R
dω
2πeiωxU(ω)
[W sε
(x, k − εω
2
)−W s
ε
(x, k +
εω
2
)].
2. Corresponding to the asymptotic Wigner equation for asymmetrical transformation (2.2.5),
one has:
∂W sε
∂t+ k
∂W sε
∂x− ∂U
∂x
∂W sε
∂k= −1
εLsW s
ε , (2.2.25)
where the skew symmetric operator Ls is given by
Lsf(z, k) = k∂f
∂z− 1
i
∑µ∈L∗
eiµx/εV (µ) [f(x, k − µ/2)− f(x, k + µ/2)] .
31
3. Corresponding to (2.2.7), one has the O(
1ε
)and O(1) expansions:
LsW s0 = 0, (2.2.26a)
∂W s0
∂t+ k
∂W s0
∂x− ∂U
∂x
∂W s0
∂k= −LsW s
1 . (2.2.26b)
4. Same as in (2.2.8), one has the following symmetrical definition for Qmm
Qsmn(z, k) = Qsmn(z, µ, p)
=
∫ 2π
0
dy
2πei(p+µ)yΨm
(z − y
2, p)
Ψn
(z +
y
2, p).
They are eigenfunctions of Ls
LsQsmn = i(Em − En)Qsmn .
5. If the eigenvalues En are well separated, i.e. Em 6= En, for m 6= n, the solution to
(2.2.26a) is:
W s0 =
∑m
σsmmQsmm.
By taking the inner product with Qsmm on both sides of (2.2.26b), one obtains the same
classical Liouville equations for σsmm as in (2.2.17)
∂tσsmm + ∂pEm∂xσ
smm − ∂xU∂pσsmm = 0.
6. If some bands touch at point pc, the solution to (2.2.25) is given by:
W sε =
∑m
σmmQsmm +
∑m 6=n
σmnQsmn.
32
In the two-band case, m,n = 1, 2, then σsmn is governed by,
∂tσs +As∂xσ
s +Bs∂pσs = −BsCsσs +
iDs
εσs, (2.2.27)
where σs = (σs11 σs12 σ
s21 σ
s22 )T , then Bs = B, Cs = C and Ds = D are the same as the
asymmetric case, while As is given by
As =
∂pE112ψ12
12ψ21 0
12ψ21 ∂p
E1+E22 0 1
2ψ21
12ψ12 0 ∂p
E1+E22
12ψ12
0 12ψ12
12ψ21 ∂pE2
and ψmn, φmn are given by (2.2.19f). Noted that As = (As)†.
We call this new system obtained by the symmetric Wigner transform (2.2.27) the Liouville-S
system. Apparently the only difference from the previous one, named Liouville-A lies in the
transport matrices A 6= As. Despite that, it could be easily shown that the two share the same
weak limit as ε → 0 (see Appendix A.1 for detail). This formally suggests that the behavior
of σ11, σ22 and σs11, σs22 in the two models is about the same. This observation is confirmed
by numerical results in Section 2.3.1.
2.2.3 Main properties of the model
The system for the non-adiabatic case we derived (2.2.19) has two transport coefficients
A and B, and C and D are the two coupling terms. We show below how Berry phase get
involved in this system, its hyperbolicity, and its connection with the model for the adiabatic
case (2.2.17).
33
Berry phase
In this part we aim at explaining the role played by Berry phase in our model. It has
been known for long that the second order correction to the semi-classical trajectory should
be related to Berry phase, see physics argument in [21, 180, 197, 198, 12] and mathematical
justification in [163, 162, 57]. Our model incorporate this phenomenon in a very subtle way
as well: the Berry connection term explicitly appears in the coupling term on the right hand
side. To compute the Berry phase, we fix one particle and follow its trajectory in momentum
space. Adiabatic theory claims that the particle moves along the energy band it starts on, as
well as it is away from the crossing zone, and thus the wave function associated to this particle
should simply be Ψm(p(t)) with a to be determined phase shift. Here p(t) is the trajectory
with p(0) = pi [180]:
ψm = Ψm(pi)→ ψm(t) = eiθm(t)Ψm(p(t)), (2.2.28)
with
θm(t) = −1
ε
∫ t
0Em(p(t′))dt′ + i
∫ t
0〈Ψm(p(t′))|∂t|Ψm(p(t′))〉dt′. (2.2.29)
The first term is called dynamical phase factor and the second term is the so-called Berry
phase term. We change the variable and obtain:
γm = i
∫ t
0〈Ψm(p(t′))|∂t|Ψm(p(t′))〉dt = i
∫ p
pi
〈Ψm(p′)|∂p|Ψm(p′)〉dp′ = i
∫φmmdp′. (2.2.30)
Remark 2.2.2. Note the phase shifting here is related to but not the same thing as cm
appeared before. cm is a function on time given Ψm is static function on x. In this picture,
however, we follow the trajectory of a particle and at each time point, we only evaluate Ψm at
one p, and thus Ψm is treated as a time-dependent function. In fact, in this case, it is named
as instantaneous eigenstates.
34
It is easy to check, the phase change of the off-diagonal element of the density matrix is
determined by this Berry phase term, that is, we write out the density matrix:
ψm(p(t))ψn(p(t)) = e−i(θm(t)−θn(t))Ψm(p(t))Ψn(p(t)), (2.2.31)
and the time change of the phase shift is:
d
dt(θm − θn) = −i∂xU(φmm − φnn)− 1
ε(Em − En). (2.2.32)
This is exactly the natural frequency of the off-diagonal term appeared in the coupling term
in (2.2.19).
Remark 2.2.3. Note that if m = n, θm − θn = 0, and this piece of information is not seen.
Hyperbolicity
The semi-classical Liouville system (2.2.19) is hyperbolic. To prove that we use the following
compact form:
∂tσ +A∂xσ − Ux∂pσ = Sσ (2.2.33)
where
S = C +iD
ε,
and σ, A, C, D are defined in (2.2.19). Hyperbolicity is easily seen since:
• A = A† is a Hermitian matrix, and can be diagonalized;
• S = −S† is a skew-Hermitian matrix.
35
Connection of the two models
The two models derived for adiabatic case (2.2.17) and non-adiabatic case (2.2.19) in 2.2.1
are intrinsically connected. In spite of the fact that these two models are for two different
scenarios, the underlying methodology is the same: expand Wigner function in terms of Qmn
and derive the associated equation. In fact, (2.2.19) is obtained upon projecting W onto the
entire basis, and thus could be regarded as the full system for a general setting. When one
applies Assumption 2.2.1 onto it, and studies the behavior of the solution away from pc where
E1 and E2 are well-separated, it should provide a picture that is consistent with (2.2.17). In
fact we could formally show that the iε terms for the transition coefficients σ12 and σ21 in the
system lead to high oscillations, and weakly, as ε→ 0, the system formally goes to the classical
one (2.2.17).
To argue that the transition coefficients are weakly zero, and that they have zero influence
on σmm in the weak sense, we assume that the initial data for the transition coefficients are
all zero, and that Ux does not change sign in time for all x, take −Ux > 0 for example, then:
Case 1. If p −C0√ε for C0 = O(1), then σ12 and σ21 are of o(
√ε);
Case 2. If −C0√ε < p < C0
√ε, then σ12 and σ21 are of O(
√ε), and σ12 and σ21 are slowly
varying, i.e. ∂tσ12 O(
1ε
), and ∂tσ21 O
(1ε
);
Case 3. If p C0√ε, σ12 and σ21 are highly oscillatory with mean 0.
Apparently Case 1 and 3 naturally unveil the connection between the two models.
The justification is tedious, and we show the proof for a simpler model that maintains the
essential properties of the system in Appendix A.1.
Remark 2.2.4. The assumption σmn(t = 0) = 0 (m 6= n) is a reasonable assumption. In fact,
given arbitrary initial condition, one could check that away from pc, the weak limit of σmn is
36
always zero, as ε→ 0, as can be seen in Appendix A.1. So numerically we treat the initial data
for both σ12 and σ21 zero, given the initial velocity ∂xS0(x) away from the crossing point, i.e.
σ(0, x, p) ≈ δ (p− ∂xS0(x))
(a2
1 0 0 a22
)T, if ∂xS0(x) 6= pc. (2.2.34)
This assumption is intuitive and empirical, but it does give us some convenience in solving the
Liouville model numerically. In fact, the numerical examples provided later indeed show that
the band-to-band transition is captured very well with initial data (2.2.34).
2.2.4 Numerical method
With the models and their properties shown above, we are capable designing numerical
methods. The full system (2.2.19) is comparably much more difficult to compute than (2.2.17)
but the latter one fails to be valid in the transition zone. Thus one needs a numerical method
that takes benefits from both, so that the right transition rates could be captured with low
computational cost. To do that, we make use of the fact shown above that the two models are
intrinsically connected:
• the classical Liouville is an approximation (in weak limit) to the full Liouville system
away from the crossing point;
• σ12 and σ21 are slowly varying in a neighborhood of pc = 0, with a small amplitude before
the characteristic hitting pc, and a rapid oscillation after that.
Heuristically we propose a domain decomposition method, which is: away from pc, when the
classical Liouville equation (2.2.17) is a good approximation, we solve this set of equations, but
around pc, in the O(ε) neighbourhood, we switch back to the full model. The gain is obvious:
numerically it is much easier and more efficient to solve the classical adiabatic system, and
thus this approach saves a great amount of computational cost than solving the full system
37
everywhere. The details are the following:
Given a fixed spatial point x, the sign of −Ux determines the traveling direction of wave in
p. Assume −Ux > 0, i.e. the wave we study is right-going:
Classical regions: p < −C0√ε and p > C0
√ε: In this region, coarse mesh independent on ε
is used to solve the classical Liouville system (2.2.17), and σ12 and σ21 are set to be zero.
Semi-classical region: p ∈ B \ Classical region: Solve the full Liouville system (2.2.19).
The incoming boundary conditions for σ12 and σ21 are set to be zero, and the incoming
boundary condition for σ11 and σ22 are the inflow boundary condition. A fine mesh is
used with ∆x and ∆t much less than√ε.
In case of −Ux < 0, and the wave if left-going, boundary condition can be set up in the same
way. The domain decomposition is carried out on p space, and thus Ux can be regarded as
constant for each x.
Remark 2.2.5. Our analysis is based on the regularity of the coefficient matrix C in the Liou-
ville system. But usually, the value of C’s element can be of O(1/δ) where δ = minp |E1(p)−
E2(p)| is the minimal band gap. So C will be large if the minimal band gap δ is small, and the
numerical discretization in the semi-classical region should resolve this small parameter δ. In
the interested regime ε ∼ δ2, o(√ε) mesh is enough.
2.3 Numerical example
In this section we show some numerical examples. Section 2.3.1 is for linear external
potential. In this special case, the solution to (2.2.19) could be explicitly expressed, and thus
numerically is treated differently. In section 2.3.2 we deal with a nonlinear external potential
U(x) using the method invented above in section 2.2.4.
38
In both examples we use the Mathieu model, i.e. the periodic potential is V (z) = cos z.
The first eight Bloch eigenvalues are shown in Figure 2.3.1. Apparently, some eigenvalues get
−0.5 0 0.5−2
0
2
4
6
8
10
p
E
(a)
−0.5 0 0.5
1.4
1.6
1.8
2
2.2
2.4
2.6
2.8
3
3.2
p
E
4th band
5th band
(b)
Figure 2.3.1: The eigenvalues of the Mathieu Model, V (x) = cosx.
very close to each other around p = 0, ±0.5.
We will focus on the 4th and 5th bands1. Denote Ψ1 and Ψ2 as the Bloch functions
corresponding to the 4th and 5th bands respectively.
For comparison, we will compare the numerical results to the ones given by the original
Schrodinger equation, computed through the methods given in [105] with mesh size and time
step much smaller than ε.
2.3.1 Linear U(x)
We deal with the linear external potential in this section:
U(x) = U0 − βx.1The minimum gap between the 4th and 5th bands is 0.0247, located at p = 0. The gap is small enough so
that the quantum effect can be seen for the ε being used.
39
Then the Schrodinger equation is:
iεψεt = −ε2
2ψεxx +
[cos(xε
)+ (U0 − βx)
]ψε, (2.3.1)
with the initial data given as a wave packet along the 4th band:
ψI = a0(x) Φ1
(xε, ∂xS0(x)
)eiS0(x)/ε, with S0(x) = p∗x. (2.3.2)
Correspondingly, the Liouville-A becomes:
∂tσ + β∂pσ = Rσ, (2.3.3a)
σ(0, x, p) = σI(x, p) = |a0(x)|2δ (p− ∂xS0(x))
(1 0 0 0
)T, (2.3.3b)
with R given by:
R = −βC +iD
ε−A∂x. (2.3.4)
To numerically compute it, the difficulty is two fold. Firstly, the solution is highly oscillatory,
and typical numerical method calls for very dense mesh that generates a great amount of
computational cost; and secondly, the initial data is a delta function and numerically it can
be tricky to solve. To resolve the first issue, we rely on the fact that β is a constant and
characteristic could be explicitly expressed, following which, the main part of the numerical
integration, the solution to the equation, could be repeated used, and is thus evaluated only
once at the beginning of computation, as will be more clear later on. To better interpret
the delta function in the initial data, instead of following the classical approach: i.e. using
a Gaussian function with small variance to approximate the delta function, with the error is
related to this variance, we will adopt a singularity decomposition idea invented in [114].
The computation is done for the Liouville-A system. Liouville-S can be computed in exactly
40
the same way.
A Fourier transform based integration method
We define the Fourier transform of f(t, x, p) with respect to x as:
f(t, η, p) =
∫Re−iηxf(t, x, p)dx.
Perform this transform on Liouville-A (2.3.3a), one gets:
∂tσ + β ∂pσ =
(−βC +
iD
ε− iηA
)σ =: Rσ
In this special case, β is a constant so the characteristic line can be obtained analytically. As a
consequence one can avoid using the dense mesh that is required by high oscillation introduced
from iDε term. We take the first time step t ∈ [0,∆t] for example. Along the characteristic line
p(t) = p0 + βt, one evaluates σ and R at (t, p(t), η) and has:
dσ
dt= R σ.
The solution to this ODE system could be explicitly written down:
σ11(t) =σ11(0)−∫ t
0iη ∂pE1 σ11(t)dt
+
∫ t
0
(β(φ12σ12 − φ21σ21)− iη ψ12σ12
)dt ,
(2.3.5a)
and
σ12(t) =e∫ t0 K
ε(τ)dτ σ12(0)
+
∫ t
0e∫ ts K
ε(τ)dτ(G(s)σ11(s)−H(s)σ22(s)
)ds ,
(2.3.5b)
41
where:
Kε =i
ε(E2 − E1)− iη ∂pE2 − β(φ11 − φ22),
and
G = −βφ12 − iηψ21, H = βφ12.
For t small, one obtains an approximation to (2.3.5):
σ11(t) ≈ σ11(0)− iηt∂pE1
(p(t)
)σ11(t)
+(βφ12(0)− iηψ12(0)
) ∫ t
0σ12dt− βφ21(0)
∫ t
0σ21dt , (2.3.6a)
σ12(t) ≈ σ12(0)e∫ t0 K
ε(τ)dτ
+(G(0)σ11(0)−H(0)σ22(0)
) ∫ t
0e∫ ts K
ε(τ)dτds . (2.3.6b)
Note that this approximation introduces (∆t)2 numerical error. Plug (2.3.6b) into (2.3.6a), to
evaluate σ11(∆t) and σ12(∆t), one needs to compute:
F0 := e∫ t0 K
ε(τ)dτ ,
F1 :=
∫ ∆t
0e∫ t0 K
ε(τ)dτ dt,
F2 :=
∫ ∆t
0
∫ t
0e∫ ts K
ε(τ)dτ ds dt.
It is not a easy job to compute these three since their integrands are highly oscillatory. However,
if one chooses ∆t such that |β|∆t = ∆p, then at each time step the grid points of p(t) lie on the
characteristics, and these three could be repeated used. In fact, we compute them only once
(with a highly resolved calculation) at the very beginning. The analysis for stability could be
tricky and we do that for a simpler model problem in Appendix A.2.
42
A singularity decomposition idea
To handle the delta function in the initial condition (2.3.3b), one usually approximates
it with a Gaussian function with small variance, and numerical error was determined by the
width of the Gaussian. What is more, the Gaussian may lose its accuracy by expanding along
the evolution and the analysis can be nasty. As stated before, in some special cases, this error
could be avoided. In fact, as we can see, (2.3.3a) is linear, and the solution keeps being a
delta function if it starts as one, in p direction, and one only needs to figure out the position
of the delta function and its amplitude. This observation obviously suggests a singularity
decomposition idea, which is also used in [114]. Write the ansatz of σmn(t, x, p) as:
σ(t, x, p) = ω(t, x, p) δ(θ(t, p)
)(2.3.7)
in which:
• θ(t, p) = p− (p∗ + βt), which solves the Liouville equation
∂tθ + β ∂pθ = 0 ,
• ω satisfies the same equation as σ:
∂tω + β ∂pω = Rω . (2.3.8)
43
These can be proved by simple derivations. Formally, one has
∂σ
∂t=
∂
∂t
(ω δ(θ)
)=∂ω
∂tδ(θ) + ω δ′(θ)
∂θ
∂t
= Rω δ(θ)− β∂ω∂p
δ(θ)− βω δ′(θ)
= Rω δ(θ)− β∂ω∂p
δ(θ)− βω δ′(θ)∂θ∂p
= Rω δ(θ)− β ∂
∂p
(ω δ(θ)
)= Rσ − β∂σ
∂p.
The equalities above should be understood in the distributional sense. The decomposition
(2.3.7) enables one to solve for ω and θ separately with good (bounded) initial data |a0(x)|2
and ∂xS0(x) respectively. The equation ω satisfies is the same as the one for σ, thus the
numerical method introduced in Section 2.3.1 can be used. In the final output, one needs to
get back to σ using (2.3.7), so a discrete delta approximation is only needed at the output
time, not during time evolution.
Numerical experiments
We implement the numerical method with the following data
β = 1, p∗ = −0.25, a0(x) = exp
(−25(x− π)2
2
). (2.3.9)
To verify the accuracy, we compare our numerical methods with the one given by directly
compute the original Schrodinger equation. The output is the density, the cumulative density
function (c.d.f.) and mass in the 1st band, defined respectively by:
ρε = |ψε|2, γε =
∫ x
−∞ρε(y)dy, m1(t) =
∫|P1ψ
ε(t, x)|2dx , (2.3.10)
44
where Pn is the projection onto the nth band:
Pnψ(x) =
∫B
dp
∫R
dyψ(y)Ψn(y, p)Ψn(x, p), φ ∈ L2(R), m ∈ N .
The two integrals in (2.3.10) are calculated by the midpoint quadrature rule numerically.
Figure 2.3.2 shows the density and c.d.f. computed for the Schrodinger equation, the
Liouville-A and the Liouville-S respectively at t = 0.5. The results match quite well.
Figure 2.3.3 shows the evolution of m1 as a function of time t. One can see the total
mass on the first band jumps down at around t = 0.25, when the momentum p reaches
pc = 0, reflecting the 4th-to-5th band transition. The experiment also shows that smaller
ε gives smaller transition rate. Note that some small oscillations occur around the crossing
region. They are related to the interference phenomena, and are usually called the Stueckelberg
oscillation [38, 175, 184].
Define L1 error in the cumulative distribution function (c.d.f.) [125, 82]:
Errε(t) =
∫R
∣∣∣∣∫ x
−∞
(ρεS(t, z)− ρεL(t, z)
)dz
∣∣∣∣ dx, (2.3.11)
where ρεS and ρεL denote the density calculated by the Schrodinger equation and the Liouville
system respectively. Numerically we compute (2.3.11) using the midpoint quadrature rule.
Figure 2.3.4 shows this at time t = 0.5. As ε → 0, the Liouville system gets more accurate,
and the error decreases with the speed of O(ε).
2.3.2 A domain decomposition computation
This section shows examples with nonlinear U(x). For this general case, the p-characteristic
is no longer a straight line, and we do not have the fast solver in the previous section any more.
To numerically solve Liouville-A, we use the domain decomposition method. The convection
45
2.5 3 3.5 4 4.5 50
1
2
3
4
x
ρε
SchrodingerLiouville-ALiouville-S
2.5 3 3.5 4 4.5 50
0.2
0.4
0.6
0.8
1
x
γε
SchrodingerLiouville-ALiouville-S
(a) ε = 180
2.5 3 3.5 4 4.5 50
0.5
1
1.5
2
2.5
3
3.5
x
ρε
SchrodingerLiouville-ALiouville-S
2.5 3 3.5 4 4.5 50
0.2
0.4
0.6
0.8
1
x
γε
SchrodingerLiouville-ALiouville-S
(b) ε = 1128
2.5 3 3.5 4 4.5 50
0.5
1
1.5
2
2.5
3
3.5
x
ρε
SchrodingerLiouville-ALiouville-S
2.5 3 3.5 4 4.5 50
0.2
0.4
0.6
0.8
1
x
γε
SchrodingerLiouville-ALiouville-S
(c) ε = 1256
Figure 2.3.2: Periodic potential problem. Linear U(x). t = 0.5. The left and right column arefor the position density ρε, and c.d.f. γε respectively. The solid line, the dash line and thedotted line are the numerical solutions to the Schrodinger, the Liouville-A and the Liouville-Srespectively.
46
0 0.1 0.2 0.3 0.4 0.50
0.2
0.4
0.6
0.8
1
t
m1
Schrodinger
Liouville-A
Liouville-S
(a) ε = 2−8
0 0.1 0.2 0.3 0.4 0.50
0.2
0.4
0.6
0.8
1
t
m1
Schrodinger
Liouville-A
Liouville-S
(b) ε = 2−9
0 0.1 0.2 0.3 0.4 0.50
0.2
0.4
0.6
0.8
1
t
m1
SchrodingerLiouville-ALiouville-S
(c) ε = 2−10
0 0.1 0.2 0.3 0.4 0.5
0.4
0.5
0.6
0.7
0.8
0.9
1
t
m1
SchrodingerLiouville-ALiouville-S
(d) ε = 2−11
Figure 2.3.3: Periodic potential problem. Linear U(x). Time evolution of m1(t) defined in(2.3.10).
47
0 0.005 0.01 0.015 0.02 0.025 0.03 0.0350
0.005
0.01
0.015
0.02
0.025
ε
Errε
Liouville-A
Liouville-S
Figure 2.3.4: Periodic potential problem. Linear U(x). Errε as function of ε at t = 0.5.
terms are handled by classical finite volume method.
We compute the Liouville-A system with both a pure and a mixed state initial data with:
U(x) = −x− sinx
2.
Example 1: A pure state initial data
In this example, we use the same pure state initial data as in the previous example
(2.3.2),(2.3.9). Correspondingly, the initial data for the Liouville-A system is given by (2.3.3b).
Numerically a Gaussian function centered at p∗ with variance of ε/16 is used to approximate
the δ−function.
Figures 2.3.5 and 2.3.6 show the density, the c.d.f at t = 0.5 and evolution of m1 (2.3.10)
computed for both the Schrodinger equation and the Liouville-A system. The numerical results
for the two systems agree well. Figure 2.3.7 gives decay of Errε (2.3.11) with respect to ε.
48
2.5 3 3.5 4 4.5 50
1
2
3
4
x
ρε
SchrodingerLiouville-A
2.5 3 3.5 4 4.5 50
0.2
0.4
0.6
0.8
1
1.2
1.4
x
γε
SchrodingerLiouville-A
(a) ε = 180
2.5 3 3.5 4 4.5 50
1
2
3
4
x
ρε
SchrodingerLiouville-A
2.5 3 3.5 4 4.5 50
0.2
0.4
0.6
0.8
1
1.2
1.4
x
γε
SchrodingerLiouville-A
(b) ε = 1128
2.5 3 3.5 4 4.5 50
0.5
1
1.5
2
2.5
3
3.5
x
ρε
SchrodingerLiouville-A
2.5 3 3.5 4 4.5 50
0.2
0.4
0.6
0.8
1
1.2
1.4
x
γε
SchrodingerLiouville-A
(c) ε = 1256
Figure 2.3.5: Periodic potential problem. Example 1. t = 0.5. The left and right columnsshow the position density ρε, and the c.d.f. γε respectively.
49
0 0.1 0.2 0.3 0.4 0.50
0.2
0.4
0.6
0.8
1
1.2
t
m1
SchrodingerLiouville-A
(a) ε = 2−8
0 0.1 0.2 0.3 0.4 0.50
0.2
0.4
0.6
0.8
1
1.2
t
m1
SchrodingerLiouville-A
(b) ε = 2−9
0 0.1 0.2 0.3 0.4 0.50.2
0.4
0.6
0.8
1
1.2
t
m1
SchrodingerLiouville-A
(c) ε = 2−10
0 0.1 0.2 0.3 0.4 0.50.4
0.5
0.6
0.7
0.8
0.9
1
1.1
t
m1
SchrodingerLiouville-A
(d) ε = 2−11
Figure 2.3.6: Periodic potential problem. Example 1. Time evolution of m1(t) defined in(2.3.10).
50
The numerical results show that the hybrid model can capture the band-to-band transition
0.014 0.016 0.018 0.02 0.022 0.024 0.026 0.028 0.03 0.0320.004
0.006
0.008
0.01
0.012
0.014
0.016
ε
Errε
Figure 2.3.7: Periodic potential problem. Example 1. The L1 error Errε between Liouville-Aand the Schrodinger solution at t = 0.5.
phenomena, and the error decays like O(ε).
Remark 2.3.1. As Ux varies with x, the wave packet becomes de-coherent. This weakens
the interference phenomenon[38, 175, 184]. As one can see in Figure 2.3.6, the Stueckelberg
oscillations around the crossing region is much weaker than those in the previous example.
Example 2: A mixed state initial data.
This example is for the case when the initial data is a mixed state:
ψI = a0(x)[Φ1
(xε, p∗)
eip∗x/ε + Φ2
(xε, p∗)
eip∗x/ε], p∗ = −0.25.
51
Correspondingly, the initial data for the semi-classical Liouville system should be:
σ = a20(x) δ(p− p∗) [1, 1, 1, 1]T .
Since p∗ is away from the crossing point and σ12 and σ21 weakly converge to zero as ε → 0,
numerically, we regard them as zero and use
σ ≈ a20(x) δG(p− p∗) [1, 0, 0, 1]T
as the initial condition, where δG(p) is a Gaussian function centered at zero.
The density and the c.d.f. are computed for the Liouville-A and the Schrodinger, compared
in Figure 2.3.8. Errε as a function of ε is shown in Figure 2.3.9.
52
1 2 3 4 50
0.5
1
1.5
2
2.5
x
ρε
SchrodingerLiouville-A
1 2 3 4 50
0.2
0.4
0.6
0.8
1
x
γε
SchrodingerLiouville-A
(a) ε = 180
1 2 3 4 50
0.5
1
1.5
2
2.5
x
ρε
SchrodingerLiouville-A
1 2 3 4 50
0.2
0.4
0.6
0.8
1
x
γε
SchrodingerLiouville-A
(b) ε = 1128
1 2 3 4 50
0.5
1
1.5
2
2.5
x
ρε
SchrodingerLiouville-A
1 2 3 4 50
0.2
0.4
0.6
0.8
1
x
γε
SchrodingerLiouville-A
(c) ε = 1256
Figure 2.3.8: Periodic potential problem. Example 2. At time t = 0.5. The left and the rightcolumns show the position density ρε, and the c.d.f. γε respectively.
53
0.012 0.014 0.016 0.018 0.02 0.022 0.024 0.026 0.028 0.030.008
0.01
0.012
0.014
0.016
0.018
0.02
ε
Errε
Figure 2.3.9: Periodic potential problem. Example 2. The L1 error Errε as a function of ε att = 0.5.
54
Chapter 3
Semi-classical description of
molecular dynamics
In this chapter we study a many body problem with N nuclei and n electrons. Semi-
classically, nuclei could jump across different energy surfaces. Around the intersection point
where two or more energy surfaces get close or even touch, such quantum transitions can be
influential. This quantum molecular dynamics problem plays an important role in analyzing
chemical reaction system, and has many interesting applications in chemistry, physics and
biology.
An interesting story is about bioluminescence of fireflies [179]. In this luciferin-luciferase
system, the luciferase-bound dioxetanone climbs up to an excited energy state through inter-
section point, producing excited-state luciferase-bound oxyluciferin, who falls back later on to
the electronic ground state, with the emission of a visible photon. Most other bioluminescence
systems undergo the same mechanism, and similar schematic models are used to interpret
them [195]. In general, most chemical reaction could be explained by such energy transition in
the quantum regime. Many other applications can also be found in [157, 158, 203]. Also see a
review in [28], and [196].
Despite its importance, the understanding to most such problems is very limited. Many
models in interpreting the real chemical reaction process are very much simplified to be prac-
tical, but even for the simplified model, very little quantitative analysis is achieved. The
55
numerical developments for this surface hopping problem around intersection points are far
from being satisfactory either. Usually to deal with the many body problems in which two
types of particles have disparate masses, people use the Born-Oppenheimer approximation.
The idea is that the heavy particles are much more reluctant to move than the light ones and
thus one could separate the motion of the two types of particles. In our problem to be specific,
energy band structure for the electrons is solved first with fixed configuration of nuclei as they
are considered at rest. However, this approximation is only valid when energy surfaces are very
well-separated, and when they get close to each other, the transition between eigenfunctions
should be taken care of.
In chemistry community, direct computation on carefully designed orthogonal basis has
been used, namely discrete variable representation. The algorithm dates back to [51, 99] and
was popularized in a series of work in 80s, see a review paper in [145]. However, the orthogonal
basis for the wave function to is usually defined on the grid points in the physical domain,
leading to high computational cost, even after fast solvers are applied [16]. Another approach
is the probabilistic algorithms that rely on the Landau-Zener formula, derived in [131, 202],
and proved with better accuracy in [91, 13, 108, 127]. The transition rate given by the formula
presents the total probability of a particle jumping from one band to another, whose trajectory
passes through the entire domain. With some simplification, it is shown that in the leading
order, the rate only depends on the smallest gap between the energy bands. This approach is
adopted in computation, called surface hopping method, initiated in [185] and studied more
thoroughly in [184, 189, 136] on the physical space, as summarized in [55]. In [130, 136]
the authors implemented the same idea in the phase space using particle method, under the
framework of the Wigner transform, and the underlying analysis is carried out in [137, 62].
Eulerian framework was adopted later on in [118, 117] for uniform global accuracy. However,
in all of these methods, it is assumed that the transition takes place locally in the spatial space
56
– around the intersection point, based on the argument that the transition away from it is
exponentially small and thus is treated as zero. Numerically it simplifies the computation a
great deal, however, due to the brutal physical simplification, part of the transition is for sure
missing.
Our approach is to look for semi-classical models in phase space that capture the correct
transition rate. We are going to derive a set of equations in phase space using Wigner trans-
formation and Weyl quantization. These two counterparts link pseudo-operators with phase
space functions in different directions. In our approach, we do not cumulate all the transi-
tions into one number, but instead the transition rates are functions in phase space, that are
naturally characterized by the off diagonal entries in the Wigner matrix. We find the equa-
tions that govern their evolution in time, and we also compare our model with the classical
model for adiabatic system where bands are assumed to be well-separated. In the adiabatic
regions, these two models are consistent. Numerically we use domain decomposition idea, and
take advantage of both models: based on the degeneracy of eigenvalues, we apply appropriate
models in different regions, so that the model could be computed cheaply with the transition
rate captured correctly.
In section 3.1 we give a brief review of the background of the problem. Our model will
be presented in section 3.2: in subsection 3.2.1 we introduce the Weyl quantization and its
associated operations including Moyal product, that will be used to derive the equations for
both adiabatic and non-adiabatic cases in subsection 3.2.2. The connection between the two
systems will be shown in subsection 3.2.3. Based on the conclusion obtained there, numerical
method are constructed. Numerical examples will be shown in section 3.3.
57
3.1 Born-Oppenheimer approximation and problem background
The Schrodinger equation for the quantum molecular dynamics is given by:
i~∂tΦ(t,x,y) = HΦ(t,x,y), (3.1.1)
Φ(0,x,y) = Φ0(x,y). (3.1.2)
Here ~ is the Planck constant, Φ is the wave function that depends on time t, position of nuclei
x ∈ R3N and position of electrons y ∈ R3n with N and n standing for the number of particles
and 3 is the dimension. H is the Hamiltonian, and in this circumstance, it is given by:
H = −N∑j=1
~2
2Mj∆xj −
n∑j=1
~2
2mj∆yj +
∑j<k
1
|yj − yk|+∑j<k
ZjZk|xj − xk|
−N∑j=1
n∑k=1
Zj|xj − yk|
.
The first two terms represent kinetic energy for the two particles, and the last three are the
Coulomb potential. mi, Mi and Zi stand for mass of electron, nucleus and electric charge
respectively. We combine the last four terms all together and call it electronic Hamiltonian,
denoted by He.
This equation is a general setting for describing chemical systems, but is cursed by its high
dimensionality for practical use. For instance the benzene molecule consists of 12 nuclei and
42 electrons, and thus has 162 spacial variable – generating very high computation cost. A
standard way to reduce the degree of freedom is to use Born-Oppenheimer approximation [22].
The approximation was derived for the Schrodinger equation for molecular dynamics and was
used to describe chemical reactions [148]. We explain in detail about the set up of the problem
in 3.1.1 and discuss its basic properties in 3.1.2.
58
3.1.1 Equation set up
The Born-Oppenheimer approximation is based on the observation that the mass ratio
between nucleus and electron is very high, and intuitively the motion of the nuclei should be
much slower than that of electrons. As argued in the original paper [22], the electrons quickly
adjust themselves in response to the slow motion of nuclei, and they stay on the n-th energy
band if they start on it, determined as though the nuclei were not moving. The impact from
the electrons to the nuclei is taken as an effective potential, and due to the disparity of the
masses, it is seen that the nuclei obey semi-classical motion. This property makes it possible
for us to separate the two motions and solve them one by one. Mathematically, one could break
the wavefunction of a molecule into the electronic and nuclear component, i.e. we write the
solution to the original Schrodinger equation as a separable function with one element dealing
with electrons and the other dealing with nuclei. We solve the part for electron first. Since
nuclei move at much slower speed, in a short time period, their configurations are regarded
fixed, and the spectrum for the electronic part with these fixed parameters are solved. In this
step, only the spatial variables of electrons are involved. The second step is to solve the nuclei
part with the spectrum, in which only nuclei are computed. In both of these two steps, only
a portion of variables are involved, which highly reduced the degree of freedom.
The whole idea starts with the marvelous intuition of Born and Oppenheimer. In their
original paper, it is assumed that the mass ratio is going to infinity, so asymptotic expansion
could be formally carried out. In the leading order, the nuclei are at their optimal configuration,
namely, the minimal point of the given energy surface. The original paper was done using
Taylor expansion assuming that the energy surfaces are smooth enough. The understanding
that the vibration and rotation could be properly separated is achieved in [58]. The rigorous
mathematical verification on its validity was not carried out till about half a century later [41,
90, 178]. For Gaussian wave packet, Hagedorn proved the validity of this approximation to high
59
orders in [93, 94], for smooth potential and the more realistic Coulomb potential respectively,
and the exponentially small error is acquired in [97]. All the analysis is carried out under the
assumption that energy surfaces are very well separated.
We start with nondimensionalization, and analyze the validity of the approximation.
Non-dimensionalization
Denote L, T and Ω as the typical scale for length, time and mass, and we rescale the
parameters:
t =t
T, x =
x
L, y =
y
L, m =
m
Ω, M =
M
Ω, (3.1.3)
The Hamiltonian is nondimensionalized too:
He(y, x) =mT 2
ΩL2He(y,x). (3.1.4)
Dropping the tildes the equation becomes:
iδm∂tΦ = −N∑j=1
δ2m
2M∆xjΦ +HeΦ, (3.1.5)
with δ = ~TΩL2 . In the scale when δ = 1 and t = t√
mM, we define ε =
√mM and have the
following:
iε∂tΦ = −N∑j=1
ε2
2∆xjΦ +HeΦ. (3.1.6)
In the equation (3.1.6), He is in the electronic Hamiltonian. As argued above, one has two
steps in the computation:
1st step: We solve for the spectrum of He in term of yj , with xj regarded as fixed parameters by
60
computing the following eigenvalue problem:
He(y,x)φk(y; x) = Ek(x)φk(y; x), k = 1, 2, · · · . (3.1.7)
Ek are called energy surfaces, and they depend on x. Here we assume He has complete
set of orthonormal eigenfunctions φk:
〈φi, φj〉 =
∫φ∗iφjdy = δij , (3.1.8)
with δij being the Kronecker delta and “∗” is the conjugate transpose. Since φi gives a
basis we expand Φ on it, and the projection coefficients, denoted by χi, depend on time
t and x:
Φ(t, x, y) =∑k
χk(t,x)φk(y; x). (3.1.9)
2nd step: Solve for χk by plugging the ansatz (3.1.9) into (3.1.6) and perform 〈·, φk〉. The equation
we get is:
iεχk =
−∑j
ε2
2∆xj + Ek(x)
χk +∑l
Cklχl, (3.1.10)
with
Ckl = 〈φk,−∑j
ε2
2∆xjφl〉y −
∑j
ε2〈φk,∇xjφl〉 · ∇xj .
In a compact form, it is equivalent to:
iεχ =
[−ε
2
2(∇x +Ageo)
2 + E
]χ, (3.1.11)
where Ageo is a matrix with its ij-th component being a Berry connection 〈Φi|∇xΦj〉. E
is a diagonal matrix with Eij = Ei(x)δij .
61
The Born-Oppenheimer approximation
In equation (3.1.10), it is the C term that complicates the system (also the Ageo term in
(3.1.11)). Depending on whether it is negligible, the Born-Oppenheimer approximation could
be valid or not, and whenever it breaks down, we perform the diabatic representation, to be
specific:
• It is shown that when energy surfaces Ek are very well separated, i.e. Ei − Ej =
O(1), ∀i 6= j, Ageo is considerably negligible compared to that of ∇x. We throw it
away, and (3.1.11) becomes the Born-Oppenheimer approximation:
iεχ =
(−ε
2
2∆x + E
)χ, (3.1.12)
• When two adjacent energy bands get close to each other, for example Ej−Ej−1 O(1),
Ageo can be as large as 1|Ej−Ej−1| . In this case, (3.1.12) no longer holds. However,
mathematically (3.1.12) has a nicer form, and we seek for similar expression by performing
diabetic presentation [2] as below.
Define a new basis by a unitary transformation Φk(t,x) = S(x)Φk(t,x). Upon this new
basis, the equation becomes:
iε∂tχ =
[−ε
2
2
(∇x + Ageo
)2+ E
]χ, (3.1.13)
where Ageo = S†AgeoS + S†(∇xS) and E = S†ES is no longer diagonal, but is still
symmetric. Select S in a smart way to make Ageo = 0 (for possibilities of such operations,
see [152]):
iεχ =
[−ε
2
2∆x + E(x)
]χ. (3.1.14)
Note E is no loner diagonalized.
62
In both of these two cases, we arrive at a similar expression with E as a function of x, either
diagonal (when BO approximation is valid) or not (when it breaks down).
For both simplicity and consistency, we adjust the notation as following:
iε∂ψε
∂t(t,x) =
(−ε
2
2∆x + V (x)
)ψε(t,x), (t,x) ∈
(R+,Rd
)(3.1.15)
ψε(0,x) = ψε0(x) (3.1.16)
Here ψ is a vector and V (x) is a symmetric matrix. Assume we are studying a 2-bands system
(with easy extension to multi-bands) and the potential reads as:
V (x) =1
2trV (x) +
u(x) v(x)
v(x) −u(x)
. (3.1.17)
The trace part is denoted by U(x) and the remainder as V (x). ψε ∈ C2 in this case.
3.1.2 Basic properties
Apparently, the Hamiltonian operator of the system is:
H =
(−ε
2
2∆x + U(x)
)I + V (x), (3.1.18)
with I as the 2× 2 identity matrix. Diagonalize the potential matrix V (x), one has:
ΘVΘ† = ΛV , (3.1.19)
63
with:
Θ(x) =1√
2
(1 + u(x)√
u(x)2+v(x)2
) 1 + u(x)√
u(x)2+v(x)2
v(x)√u(x)2+v(x)2
− v(x)√u(x)2+v(x)2
1 + u(x)√u(x)2+v(x)2
. (3.1.20)
Here:
ΛV = diag(λ+V , λ
−V
)= diag
(√u2(x) + v2(x),−
√u2(x) + v2(x)
), (3.1.21)
is a diagonal matrix with the two eigenvalues denoted as λ±V . We also use χ± to denote the
two columns of Θ† standing for normalized eigenvectors associated with λ±V .
If the two eigenvalues are very well separated for all x, one could diagonalize the system
and compute two separate Schrodinger equations. However, sometimes the two eigenvalues
degenerate, as u(x) = v(x) = 0 here; and when this happens, the eigenvector space associated
with the single eigenvalue is 2 dimensional, and one could not distinguish two quantum states.
In fact, ambiguity emerges in defining the right eigenfunctions: limit for (3.1.20) does not even
exist as u(x) and v(x) go to zero. The standard mathematical decomposition method breaks
down at the surface hopping point, and it is our main goal to find the right mathematical
description of the system that keeps its validation even at the intersection.
3.2 Semi-classical models
In this section we use the Wigner transform approach to derive a semi-classical model,
for both adiabatic and non-adiabatic cases. As defined in (1.2.1), the Wigner function is the
inverse Fourier transform of the density matrix:
F ε(x,p) =1
(2π)d
∫ρε(x− εy
2,x +
εy
2
)eip·y dy, (3.2.1)
64
where ρε is the density matrix:
ρε(x,x′) = ψε(x)⊗ ψε(x′). (3.2.2)
ψε is the complex conjugate of ψε. As in (1.2.2), one could express the formula in bra-ket
language [201]:
F ε(x,p) =
∫Rd
dy
(2π)d〈x− ε
2y|ρ|x +
ε
2y〉eip·y, (3.2.3)
with ρ = |ψε〉〈ψε| standing for the density operator.
Unlike in the scalar case, here the Wigner function is a 2×2 matrix, and the collision term
is much more complicated than the one in (1.2.5). Simple derivation gives, if ψε solves the
Schrodinger equation (1.1.1), then F solves the Wigner equation:
∂tFε + p · ∇xF
ε + Ξ[UI + V ]F ε = 0, (3.2.4)
where Ξ[V ] is an operator that, when acted on the density matrix F ε, gives:
Ξ[V ]F ε =
∫R2d
V(x− εy
2
)F ε(x,p′)− F ε(x,p′)V
(x + εy
2
)−i(2π)dε
ei(p′−p)·ydp′ dy.
Note that unlike the usual case in (1.2.6) where V (x − εy2 ) − V (x + εy
2 ) gives ∇xV (x) in the
limiting regime, here both F and V are matrices, and it is no longer guaranteed that F and
V commute. The techniques we use to overcome this difficulty is the Weyl quantization to be
introduced in the following subsection.
3.2.1 Weyl quantization and Wigner transform
The Weyl quantization was found in [192] in 1927, and was brought into the field in [74].
It was firstly adopted in [153] for non-adiabatic situation, and later on extended to graphene
65
controlled by the Dirac equation [154]. The computation for the avoided crossing is done
in [36].
The Weyl quantization maps functions defined in phase space to pseudo-operators. As will
be shown later, the Wigner transform could be regarded as the inverse Weyl quantization.
The restrictions for the Wigner equation do not appear in the pseudo-operator space, and our
strategy is to perform diagonalization on this space and map it back. We mention here that
the way to bridge the two spaces is not unique [100] and all these mapping techniques are
equivalent as summarized in [40].
Weyl Quantization
Given one function A(x, p), called a symbol, in phase space, the Weyl quantization gives
its associated operator A =W(A) such that, for an arbitrary function h(x,y) ∈ S(Rd × Rd),
W(A)[h](x) = A[h](x) =1
(2πε)d
∫ ∫A
(x + y
2,p
)h(x,y) e
iε(x−y)·p dp dy. (3.2.5)
We list two examples here, both of which can be justified by simple derivation:
1. if A(x,p) is the Wigner function F ε, then
F [h](x) =1
(2πε)d
∫ρ(x,x′)h(x′) dx′. (3.2.6)
In bra-ket language, F is exactly ρ = |ψ〉〈ψ|, and F [h](x) is simply 〈x|ψ〉〈ψ|h〉. Therefore,
in this case, the Weyl quantization can be regarded as the inverse Wigner transform.
2. if A(x,p) is the Hamiltonian H(x,p) =(p2
2 + U(x))I + V (x), then:
H[h](x) =
(ε2
2∆x + U(x)I
)h(x) + V (x)h(x) (3.2.7)
66
is the Hamiltonian operator.
We also list a few important properties [103]:
(a) the mapping is one to one, and the inverse mapping is denoted by A =W−1(A);
(b) W(A#B) =W(A)W(B); where the operation # is called the Moyal product defined as:
A#B := Aeiε2
(←−∇x·−→∇p−
←−∇p·−→∇x
)B, (3.2.8)
where the arrows indicate on which symbol the gradients act. Another less useful form
for Moyal product is:
A#B =1
(2π)2d
∫A(x− ε
2η,p +
ε
2µ)B(x′,p′)ei(x−x′)·µ+i(p−p′)·ηdµdx′dηdp′.
Both these two formulas could be verified through simple derivation.
(c) if A is a function only on x, then A = A;
von Neumann equation
With the Schrodinger equation (3.1.15), one could derive the equation satisfied by F : the
von Neumann equation [27, 174]:
iεF = [H, F ] (3.2.9)
where [A, B] = AB − BA is the commutation operation and H is the Hamiltonian defined in
(3.1.18).
Remark 3.2.1. (3.2.9) is in fact very easy to be derived in bra-ket language. We start with
the Schrodinger equation iε∂t|ψε〉 = H|ψε〉 , and its conjugate transpose −iε∂t〈ψε| = 〈ψε|H.
67
The von Neumann equation is their simple combination:
iε∂tF = iε∂t (|ψε〉〈ψε|) = iε (∂t|ψε〉) 〈ψε|+ iε|ψε〉∂t〈ψε|
= H|ψε〉〈ψε| − |ψε〉〈ψε|H = [H, F ].
Here the symbol associated to H is not diagonalized. By the structure given from (3.1.19)
to (3.1.21), we define Θ and Θ† as the Weyl quantization of Θ and Θ† respectively and multiply
them on two sides of (3.2.9). Symbolically we get:
iε∂F ′
∂t=[H ′, F ′
]=
[−∆x
2+ U(x)I, F ′
]+[ΛV , F
′]. (3.2.10)
Here we use new density operator F ′ = Θ F Θ†, new Hamiltonian H ′ = Θ H Θ†. According to
the definition of Θ, ΘV Θ† = ΛV .
Remark 3.2.2. Note the new Hamiltonian is not diagonalized, but the potential part is. The
symbol for the new Hamiltonian is
(p2
2+ U(x)
)I + ΛV .
We will show computation for the exact expression of H ′ later on in (3.2.17).
3.2.2 Adiabatic and non-adiabatic models
In this section we make use of the tools developed above to derive the model for both the
adiabatic and non-adiabatic situations. The computation is done only formally.
68
Adiabatic case
Define the crossing set:
S = x : λ+(x) = λ−(x). (3.2.11)
In this section we deal with adiabatic case and the energy bands are assumed to be very well
separated, i.e. the following assumption holds:
Assumption 3.2.1. The eigenvalues of the system do not degenerate, i.e. the two energy
bands are very well separated. In our case, where the potential term is of the form (3.1.17), it
is simply λ−V 6= λ+V 6= 0, and u2(x) + v2(x) is never zero. In other words, S is empty.
Following [74, 147], we define the matrices:
Π±(x,k) = χ±(x,k)⊗ χ±(x,k).
They reflect the orthogonal projection onto the eigen-spaces correspond to λ±(x,k) respec-
tively. With these, we study the following limit of the Wigner function, defined as:
limε→0
F (t,x,p) = F 0(t,x,p).
According to Theorem 6.1 in [74], outside the crossing set, the Wigner measure can be decom-
posed as:
F 0(t, ·) = Π+F0(t, ·)Π+ + Π−F
0(t, ·)Π− (3.2.12a)
= f+(t, ·)Π+ + f−(t, ·)Π−, (3.2.12b)
where f± = Tr(Π±F0(t, ·)) are scalar functions standing for the phase space probability den-
sities on the associated energy levels. The first equivalence comes from the fact that F 0
69
commutes with the projection operator Π± as proved in the theorem, and (3.2.12b) is because
the eigen-spaces are one-dimensional. The equation that controls the propagation of f± are:
∂tf± +∇pλ
± · ∇xf± −∇xλ
± · ∇pf± = 0, t > 0, x ∈ Rd \ S, p ∈ Rd,
f±(t = 0,x,p) = Tr(Π±F0(t = 0,x,p)) (3.2.13)
Noted that the theory above is based on the adiabatic assumption, so that degeneracy is
avoided and one is able to distinguish the projections onto different functions.
Quantum transition in non-adiabatic case
In this subsection we investigate the situation when the assumption breaks down and energy
bands get close to each other. This situation cannot be avoided as long as the trajectories of
f± come across S. In fact, dates back to [188], people have realized that the crossing set, if ever
exists, is of measure zero, however, its influence is significant. In history, to fix this new problem
on the degeneracy, people have tried adopting Landau-Zener formula in both Lagrangian and
Eulerian approach with finite volume or Gaussian beam numerical methods [185, 136, 118, 117].
Our idea is based on the von Neumann equation in (3.2.9) [153, 154], upon which we apply
the inverse Weyl map. As one can see later, this approach naturally gives the functions that
represent the transition rate. It will also be shown that away from S, the transition rate is
weakly zero and makes no impact on the whole system. This is consistent with (3.2.13). In
this sense, we consider (3.2.13) as the weak limit of the new model.
As in (3.2.10), we have obtained a diagonalized equation on the symbolic side. To obtain
its counterpart in phase function space, we perform the inverse Weyl mapping, and the new
evolution equation reads as:
iε∂F ′
∂t=[H ′, F ′
]#, (3.2.14)
70
where the commutator [A,B]# = A#B − B#A with # being the Moyal product defined in
(3.2.8). and H ′ and F ′ are symbols associated to H ′ and F ′ respectively:
H ′(x,p) = Θ(x)#H(x,p)#Θ(x)†
F ′(x,p) = Θ(x)#F (x,p)#Θ(x)†. (3.2.15)
To further derive the equation, we truncate the Moyal product to its second order:
A#B = AB − iε
2A,B+O(ε2), (3.2.16)
where the Poisson bracket is A,B = ∇pA·∇xB−∇xA·∇pB . Thus the symbol H ′ becomes:
H ′(x,p) = Λ(x,p) + iεp · ∇xΘ(x)Θ†(x) +ε2
2∇xΘ(x) · ∇xΘ†(x). (3.2.17)
Plug it into (3.2.14), equation for F ′ is obtained, and to the first order of ε, the semi-classical
limit gives (derivations are omitted here and details could be found in Appendix A.3):
∂f+
∂t = −p · ∇xf+ +∇x
(U + Eg
)· ∇pf
+ + 2ξ<f i
∂f−
∂t = −p · ∇xf− +∇x
(U − Eg
)· ∇pf
− − 2ξ<f i
∂f i
∂t = −p · ∇xfi +∇xU · ∇pf
i + ξ(f− − f+) +2Egiε f
i
. (3.2.18)
Here < stands for the real part, and the energy gap is 2Eg = 2√u2 + v2. ξ is defined by:
ξ =1
2Egp · (∇xv(x) cosφ−∇xu(x) sinφ) , (3.2.19)
with φ being the angle of the potential: cosφ = u(x)Eg
and sinφ = v(x)Eg
. The three scalar
71
functions are the four components of F ′ as in:
F ′ =
f+ f i
f i f−
. (3.2.20)
We split f i into the real part and imaginary part by defining fR =√
22 (f i + f i) and f I =
√2
2i (f i − f i). Then the system becomes:
d
dt~f = −p · ∇x ~f +A · ∇p ~f +B · ~f, (3.2.21)
with:
A = diag(∇x(U + Eg),∇x(U − Eg),∇xU,∇xU),
B =
0 0√
2ξ 0
0 0 −√
2ξ 0
−√
2ξ√
2ξ 02Egε
0 0 −2Egε 0
.
B is anti-symmetric, and thus the evolution operator is unitary, so the L2 norm of this hy-
perbolic system remains unchanged. In this setting, f± reveal the projection coefficients onto
different bands, and f i naturally arise representing the transition rate between two bands. Our
numerical method is designed for this equation.
72
Initial condition for the Wigner function in the semi-classical limit
In this subsection we derive the initial condition in the semi-classical regime. In particular,
we concern about the transform of the following initial data for the Schrodinger system (1.1.1):
ψε(t = 0,x) = ψε0(x) = gε0(x)(a+χ+(x) + a−χ−(x)
), (3.2.22)
where, gε0 is an ε-scaled Gaussian packet:
gε0(x) =1
(πε)d/4exp
− 1
2ε|x− x0|2 +
i
εp0 · (x− x0)
. (3.2.23)
Plug it into the definition of the associated Wigner function (3.2.1), we get:
F ε(t = 0,x,p) ∼ 1
(πε)dexp
−1
ε|x− x0|2 −
1
ε|p− p0|2
×((a+)2 χ+ ⊗ χ+ + (a−)2 χ− ⊗ χ−
). (3.2.24)
and in the zero limit of ε, it goes to:
δ(x− x0,p− p0)((a+)2 χ+ ⊗ χ+ + (a−)2 χ− ⊗ χ−
), as ε→ 0.
The corresponding diagonalized symbol is F ′ defined in (3.2.15):
F ′(t = 0,x,p) = Θ(x)#F (t = 0,x,p)#Θ(x)†,
ε→0−−−→ Θ(x)F (t = 0,x,p)Θ(x)†. (3.2.25)
73
To be specific, simple calculation gives:
f+(t = 0,x,p) = (a+)2
(πε)dexp
−1ε |x− x0|2 − 1
ε |p− p0|2
f+(t = 0,x,p) = (a−)2
(πε)dexp
−1ε |x− x0|2 − 1
ε |p− p0|2
fR(t = 0,x,p) = f I(t = 0,x,p) = 0
. (3.2.26)
3.2.3 Properties and numerical methods
It is easily seen that the system is hyperbolic, with a source term that is anti-symmetric,
and thus the L2 norm of the whole system is conserved.
As in the precious chapter, we intend to design an efficient numerical method based on
the possible link between the set of equations (3.2.13) that is easier to compute and the
hard one (3.2.21). The argument is very similar to what we had in the periodic potential
case. In fact, as we will see later, (3.2.21) contains the full information, whether the Born-
Oppenheimer approximation holds or not, and away from the crossing zone S when Assump-
tion 3.2.1 holds, this full system automatically converges to (3.2.13). Numerically since the
adiabatic model (3.2.13) is much easier to be computed, and thus one should stick to it when-
ever possible. Therefore, we decompose the domain into different regions depending on the gap
between energy bands, and apply different models in different regions, in particular (3.2.21)
around S and (3.2.13) away from it. The idea was adopted from [35], and implemented in [36].
There are many types of intersection, and we focus our attention to the case of avoided cross-
ing [92].
Avoided crossing
Avoided crossing is one type of band structure that was introduced in [92]. For this type
of band structure, the energy gap is a function not solely dependent on the spacial variable
74
x, but also relies on a parameter δ ≥ 0: the energy gap E is always small but gets to zero if
and only if δ = 0. In recent years avoided crossing is extensively studied: one could find the
technical analysis on the relationship between δ and ε in [172] where it is shown that the most
interesting case would be when these two parameters are at the same order. This is to say, the
gap between the levels shrinks at the same order the nucleus increase the mass. For different
type of avoided crossings, Hagedorn and Joye were able to evaluate the propagation of the
Gaussian wave packet through avoided crossing by constructing the inner and outer domain
and asymptotically match the results on two sides [95, 96]. More general initial data were
considered in [62, 61, 60] in the Wigner space. See also a review on both theory and numerics
in [24].
One typical example for avoided crossing is given in 2D. We write x = (x, y) and p = (p, q),
and use the following potential:
u(x, δ) = x, v(x, δ) =√y2 + δ2. (3.2.27)
The trace U(x) is set to be 0. So the equation (3.2.21) becomes:
∂f+
∂t = −p · ∇xf+ +∇xEg · ∇pf
+ +√
2 ξ fR
∂f−
∂t = −p · ∇xf− −∇xEg · ∇pf
− −√
2 ξ fR
∂fR
∂t = −p · ∇xfR + 2
εEgfI +√
2 ξ (f− − f+)
∂fI
∂t = −p · ∇xfI − 2
εEgfR
(3.2.28)
where in this case:
Eg =√x2 + y2 + δ2, ξ =
1
2(x2 + y2 + δ2)(qx
y√y2 + δ2
− p√y2 + δ2).
75
The potential matrix is marked as V (x, δ) with its two eigenvalues:
λ±(x,p, δ) =p2
2± Eg, (3.2.29)
With zero δ, it becomes conical crossing with the two eigenvalues touching at the origin, i.e.
λ+(xc,p, δ) = λ−(xc,p, δ) when and only when δ = 0. xc represents the crossing point, and is
the origin in this case.
Weak limit and domain decomposition
Numerically, the main difficulty comes from the high oscillations introduced by the small
ε. When ε is extremely small, with O(1)√x2 + y2, fR and f I present very rapid oscillations
with the wavelength at the scale of ε. However, one could check, that when E is nonzero, both
of these two oscillatory functions fR and f I have zero weak limit as ε→ 0, and formally, the
system goes to the adiabatic formulation:
∂f+
∂t = −p · ∇xf+ +∇xEg · ∇pf
+,
∂f−
∂t = −p · ∇xf− −∇xEg · ∇pf
−.
(3.2.30)
The situation here is very similar to the periodic case, where we adopt the domain decom-
position idea and solve appropriate set of equation in different region: in the regions where
the oscillation is strong, we compute the weak limit (3.2.13) as a good approximation, and
in the regions where the function varies slowly, we stick to the original equation (3.2.28).
Two main problems remain to be resolved: one is where to set up the interface and how to
give the boundary condition. They will be answered after the following argument about the
convergence.
Take p > 0 and x0 < 0 for example:
76
Case 1. If Eg C0√ε before hitting the crossing zone, then fR and f I are of o(
√ε);
Case 2. If −C0√ε < Eg < C0
√ε, then fR and f I are of O(
√ε), and fR and f I are slowly
varying, i.e. ∂tfi O
(1ε
);
Case 3. If Eg C0√ε after hitting the crossing zone, fR and f I are highly oscillatory with
mean 0.
As such, we propose the following decomposition:
Classical regions: Eg > C0√ε: In this region, a coarse mesh independent of ε is used to
solve the adiabatic Liouville system (3.2.30). fR and f I are set to be zero.
Semi-classical region: Eg ≤ C0√ε: Solve (3.2.28). The incoming boundary conditions for
fR and f I are set to be zero, and the incoming boundary condition for f+ and f− are
the inflow boundary condition. A fine mesh is used with ∆x and ∆t much less than√ε.
3.3 Numerical examples
In this section we show several numerical results and compare with the solution to the
original Schrodinger equation. Let ψε(t,x) be the solution of (3.1.15) with initial data given
by (3.2.22). For the lower(-) and upper(+) bands, we define the position density ρ±schr and the
population P±schr:
ρ±schr(t,x) = |Π±ψε(t,x)|2, and P±schr =
∫Rdρ±schr(t,x) dx. (3.3.1)
Correspondingly, we recover these quantities from the semi-classical Liouville system (3.2.21)
by
ρ±liou(t,x) =
∫Rdf±(t,x,k)dk, and P±liou =
∫Rdρ±liou(t,x) dx. (3.3.2)
77
Define L1 error in the cumulative distribution function (c.d.f.) [82, 125]:
Errε =
∫Rd
∣∣∣∣∫Ωx
(ρ+schr(y)− ρ+
liou(y))dy
∣∣∣∣+
∣∣∣∣∫Ωx
(ρ−schr(y)− ρ−liou(y)
)dy
∣∣∣∣dx,
where, Ωx =y ∈ Rd : yi ≤ xi, i = 1, · · · , d
for x = (x1, · · · , xd) ∈ Rd.
3.3.1 Example 1: 1D, pure state initial data
In this example, we deal with the problem in 1D with a pure initial data. Set u and v as
u(x) = x, and v(x) ≡ δ =
√ε
16. (3.3.3)
δ is of O(√ε), larger than 0 and is independent of the spatial variable x. The difference
between the two energy bands is ∆E = 2E that gets its minimum 2δ at x = 0. The initial
data for the Schrodinger equation is given in (3.2.22-3.2.23) with a+ = 1, a− = 0, x0 = 5√ε,
and p0 = −1. For comparison, we compute both the Schrodinger equation as a reference, and
the Liouville system. The Schrodinger equation is computed using the classical time-splitting
spectral method, with ∆x = ε/8 and ∆t = 5ε32 and Liouville semi-classical system is done
through the domain decomposition method. In Figure 3.3.1 we compare the results ρ± given
by the two systems, and in Figure 3.3.2 we check the evolution of the population on the first
band P+ along the time. In Figure 3.3.3, it is showed that the domain decomposition modeling
error (3.3) decreases as O(ε).
3.3.2 Example 2: 1D, mixed state initial data
In this example, we have mixed states as the initial data in one dimension. u and v are
defined in the same way as in (3.3.4). Initial wave packet for the Schrodinger equation is given
in (3.2.22-3.2.23) with a+ = a− = 1/√
2, x0 = 5√ε, and p0 = −1, thus is mixed states. In
78
Figure 3.3.4 we compare the numerical results on ρ± and P+ of the Schrodinger equation and
the Liouville system, and in Figure 3.3.5, the cumulative error is given as a function of ε. The
numerical results show O(ε) decay in error.
3.3.3 Long time behavior
In this subsection we consider the evolution that is long enough so that the data goes
through two surface hopping processes. The numerical results are provided for both the original
Schrodinger equation and its semi-classical Liouville system. Initial data is a Gaussian packet
as in (3.2.22)-(3.2.23) with a+ = 1, a− = 0, x0 = 0.3125, and p0 = −1. ε is set as 2−10. In
Figure 3.3.6 we show the evolution of P+ with respect to time. Initially, the profile is located
to the right of the crossing point, the origin in this case, with negative average velocity, it
goes to the origin and around t = 0.25 when the center of the packet roughly hits the crossing
point, more than half of the portion jump over to the lower energy level, so that we have two
packets along the two separate bands after that, and both of them keep moving towards left.
The packet on the lower energy level keeps accelerating as the potential diminishes, so it never
comes back. But the one on the higher energy level gradually decelerates since its potential
gets higher and higher, and at some point, the velocity decreases to zero and the packet starts
bouncing back towards the origin. Around t = 0.275, it gets to the crossing point for the
second time, and undergoes another hopping process. The results for the two systems match
quite well.
3.3.4 Example 4: 2D pure initial data
In this example, we deal with the problem in 2D with a pure initial data. Set u and v as
u(x) = x, and v(x) =√y2 + δ2. (3.3.4)
79
δ is chosen at√ε/2, larger than 0 and is independent of the spatial variable x. The difference
between the two energy bands is ∆E = 2E that gets its minimum 2δ at x = y = 0. The initial
data for the Schrodinger equation is given in (3.2.22-3.2.23) with a+ = 1, a− = 0, x0 = 5√ε,
y0 = 0, p0 = −1, and q0 = 0. For comparison, we compute both the Schrodinger equation
as a reference, and the Liouville system. The Schrodinger equation is computed using the
classical time-splitting spectral method, with ∆x = ∆y = ε/8 and ∆t = 5ε32 and Liouville
semi-classical system is done through the domain decomposition method with ∆x = ∆p = h
in the classical regions and ∆x = ∆p = h/2 in then semi-classical region, where h = O(√ε).
In Figure 3.3.7 we compare the results ρ± given by the two systems, and in Figure 3.3.8 we
check the evolution of the population on the first band P+ along the time. In Figure 3.3.9, it
is shown that the domain decomposition modeling error (3.3) decreases as O(√ε).
80
−0.4 −0.35 −0.3 −0.25 −0.2 −0.15 −0.1−0.5
0
0.5
1
1.5
2
2.5
3
x
ρ+
SchrodingerLiouville
−0.4 −0.35 −0.3 −0.25 −0.2 −0.15 −0.10
1
2
3
4
5
6
7
8
9
10
x
ρ-
SchrodingerLiouville
(a) ε = 2−9
−0.3 −0.25 −0.2 −0.15 −0.1−0.5
0
0.5
1
1.5
2
2.5
3
3.5
4
x
ρ+
SchrodingerLiouville
−0.3 −0.25 −0.2 −0.15 −0.10
5
10
15
x
ρ-
SchrodingerLiouville
(b) ε = 2−10
−0.22 −0.2 −0.18 −0.16 −0.14 −0.12 −0.1 −0.08 −0.060
1
2
3
4
5
6
x
ρ+
SchrodingerLiouville
−0.22 −0.2 −0.18 −0.16 −0.14 −0.12 −0.1 −0.08 −0.060
5
10
15
20
25
x
ρ-
SchrodingerLiouville
(c) ε = 2−11
Figure 3.3.1: Surface hopping problem. Example 1. The left column is for ρ+schr/liou, the
density on the upper band, and the right column gives the results of ρ−schr/liou. ∆ =√ε/2.
81
0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.80
0.1
0.2
0.3
0.4
0.5
0.6
0.7
0.8
0.9
1
t
Population
SchrodingerLiouville
(a) ε = 2−7
0 0.1 0.2 0.3 0.4 0.5 0.60
0.1
0.2
0.3
0.4
0.5
0.6
0.7
0.8
0.9
1
t
Population
SchrodingerLiouville
(b) ε = 2−8
0 0.05 0.1 0.15 0.2 0.25 0.3 0.35 0.40
0.1
0.2
0.3
0.4
0.5
0.6
0.7
0.8
0.9
1
t
Population
SchrodingerLiouville
(c) ε = 2−9
0 0.05 0.1 0.15 0.2 0.25 0.30
0.1
0.2
0.3
0.4
0.5
0.6
0.7
0.8
0.9
1
t
Population
SchrodingerLiouville
(d) ε = 2−10
0 0.05 0.1 0.15 0.20
0.1
0.2
0.3
0.4
0.5
0.6
0.7
0.8
0.9
1
t
Population
SchrodingerLiouville
(e) ε = 2−11
0 0.05 0.1 0.150
0.1
0.2
0.3
0.4
0.5
0.6
0.7
0.8
0.9
1
t
Population
SchrodingerLiouville
(f) ε = 2−12
Figure 3.3.2: Surface hopping problem. Example 1. Time evolution of the population (definedin (3.3.1) and (3.3.2)) on the upper band P+
schr/liou. ∆ =√ε/2.
82
−10 −9 −8 −7 −6−11.5
−11
−10.5
−10
−9.5
−9
−8.5
−8
−7.5
−7
log2(ε)
log 2(E
rrε)
δ =√ε/4
δ =√ε/8
δ =√ε/16
Figure 3.3.3: Surface hopping problem. Example 1. Errε decreases at O(ε). t = 10√ε.
83
−0.5 −0.4 −0.3 −0.2 −0.1 00
0.5
1
1.5
2
2.5
3
3.5
4
4.5
x
ρ+
Schrodinger
Liouville
−0.5 −0.4 −0.3 −0.2 −0.1 00
1
2
3
4
5
6
x
ρ-
SchrodingerLiouville
(a) ε = 2−9
−0.25 −0.2 −0.15 −0.1 −0.05 00
1
2
3
4
5
6
7
x
ρ+
Schrodinger
Liouville
−0.25 −0.2 −0.15 −0.1 −0.05 00
1
2
3
4
5
6
7
8
x
ρ-
SchrodingerLiouville
(b) ε = 2−10
−0.2 −0.15 −0.1 −0.05 00
2
4
6
8
10
12
14
x
ρ+
Schrodinger
Liouville
−0.2 −0.15 −0.1 −0.05 00
5
10
15
x
ρ-
SchrodingerLiouville
(c) ε = 2−11
Figure 3.3.4: Surface hopping problem. Example 2. ∆ =√ε/2. The left column is the density
on the upper band ρ+schr/liou and the right column is ρ−schr/liou.
84
−12 −11 −10 −9 −8 −7 −6−11
−10
−9
−8
−7
−6
−5
−4
log2(ε)
log 2(E
rrε)
δ =√ε/4
δ =√ε/8
δ =√ε/16
Figure 3.3.5: Surface hopping problem. Example 2. Errε as a function of ε at t = 10√ε.
0 0.5 1 1.5 2 2.5 3
0.2
0.4
0.6
0.8
1
t
Population
SchrodingerLiouville
Figure 3.3.6: Surface hopping problem. Example 3. Long time behavior of P+.
85
(a) Schrodinger Solutions
(b) Liouville Solutions
Figure 3.3.7: Surface hopping problem. Example 4. The left column is for ρ+schr/liou, the
density on the upper band, and the right column gives the results of ρ−schr/liou. δ =√ε/2.
86
0 0.2 0.4 0.6 0.8 1 1.2
0.6
0.65
0.7
0.75
0.8
0.85
0.9
0.95
1
1.05
t
Population
SchrodingerLiouville-1Liouville-2Liouville-3Liouville-4
(a) ε = 2−6
0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8
0.6
0.65
0.7
0.75
0.8
0.85
0.9
0.95
1
1.05
t
Population
SchrodingerLiouville-1Liouville-2Liouville-3Liouville-4
(b) ε = 2−7
0 0.1 0.2 0.3 0.4 0.5 0.6
0.65
0.7
0.75
0.8
0.85
0.9
0.95
1
1.05
t
Population
SchrodingerLiouville-1Liouville-2Liouville-3Liouville-4
(c) ε = 2−8
0 0.05 0.1 0.15 0.2 0.25 0.3 0.35 0.4
0.65
0.7
0.75
0.8
0.85
0.9
0.95
1
1.05
t
Population
SchrodingerLiouville-1Liouville-2Liouville-3Liouville-4
(d) ε = 2−9
Figure 3.3.8: Surface hopping problem. Example 4. Time evolution of the population onthe upper band P+
schr/liou. δ =√ε/2. The legend “Schrodinger” represents the solution of
the Schrodinger equation, “Liouville-j” represents the solution of the Liouville system by thedomain decomposition method with ∆x = ∆p = h in the classical regions and ∆x = ∆p = h/2in then semi-classical region, where h =
√ε/2j−1, and j = 1, 2, 3, 4
87
−10 −9 −8 −7 −6
−7
−6.5
−6
−5.5
−5
log2(ε)
log 2(E
rrε)
Figure 3.3.9: Surface hopping problem. Example 4. Errε at t = 10√ε. The Liouville system
by the domain decomposition method with ∆x = ∆p =√ε/8 in the classical regions and
∆x = ∆p =√ε/16 in then semi-classical region.
Part II
Kinetic Theory
88
89
In this part we study the hydrodynamic limit of kinetic theory. We are especially interested
in the compressible Euler limit of the Boltzmann equation, and focus on designing proper
numerical algorithms for it.
The Boltzmann equation describes the evolution of the probability density distribution
of rarefied gas. When the typical length and time scale is big enough, mathematically the
Knudsen number, a parameter that is the ratio of local mean free path and the typical domain
length, is small, and the rarefied gas system could be treated as continuum. We are interested
in finding the connection between these two levels of descriptions on the numerical level.
The efforts to connect statistical mechanics and fluid dynamics date back to Boltzmann,
who tried to develop “mathematically the limiting processes” that “lead from the atomistic
view to the laws of motion of continua”. Mathematical treatment started booming since
Hilbert marked it as his 6th problem in 1900. Since then, it has attracted a great amount
of attention: people study this problem out of mathematical curiosity, and out of practical
concern. Mathematically, the problem is two fold: how to study the entropy decay, and how
to link the kinetic equation to the fluid dynamics. See an early book [34] by Cercignani and
the most advanced analysis collected in [186, 173] by Villani and Saint-Raymond respectively.
On the theoretical level, the picture is rather clear. The formal connection between the
(in)compressible Euler or Navier-Stokes and the Boltzmann equation has been established, and
the rigorous proof is complete for the link to the incompressible Navier-Stokes equation. Our
target to present the same connection on the numerical level: as the Euler equations are the
hydrodynamic limit of the Boltzmann equation, we are seeking accurate numerical methods
for the Boltzmann equation that capture the solution to the Euler equations automatically in
this limit.
To this end, we introduce the concept, “asymptotic preserving”. As will be clearer later
on, the whole area explores possibilities of relaxing discretization from the strong requirement
90
of the scaling parameter. Typically, an asymptotic preserving method is a cheap solver that
preserves the asymptotic limit, on the numerical level. The idea has been very widely used in
implementing multi-scale problems, but in this thesis, we only focus our attention on one of
its many facets: which is the link between the Boltzmann equation and the compressible Euler
equations.
In chapter 4, as a introduction, we briefly review the established analytical background of
the Boltzmann equation. In chapter 5 we go over two major asymptotic preserving methods
currently available, to be specific, the BGK penalization method and the Exponential Runge-
Kutta method. Both of the two methods were initially designed for single species systems, and
we explore their extension to multi-species systems in chapter 6.
91
Chapter 4
Kinetic theory and the
hydrodynamic limit
In this chapter we briefly go over the basics of kinetic theory and show the hydrodynamic
limit of the Boltzmann equation.
4.1 Boltzmann equation
Kinetic theory aims at describing a system constituted of a large number of particles on a
mesoscopic level. We usually model the system by a distribution function f in phase space. In
case of monatomic gas, the distribution function is f = f(t, x, v), that depend on time t > 0,
position x ∈ Ω and velocity v ∈ R3, presenting the number of particles in an infinitesimal
volume dxdv around the point (x, v) at time t.
Remark 4.1.1. The Boltzmann equation is not ab initio, and its derivation, though not
within our focus in this thesis, is a remarkable thing by itself. The whole process was pioneered
by Grad in 1958 in [85], and the rigorous proof on the validity of the Boltzmann equation is
later on obtained in [132] for very short time, also see a review [133]. The general approach
is that, one starts with a collection of N particles, and traces trajectory of each particle. The
particle obeys Newton’s 2nd law, so the system is time-reversible. A distribution function
fN is a delta function that lives on 6N dimensional space R3N × R3N (3D for position and
3D for momentum for each particle in the system). f is seen as the marginal distribution
92
function of one particle of fN . From fN to f , some assumptions are to be made: N is big
enough and particles are considered identical, so that fN is sufficiently close to tensor product
of N one particle distribution function fs (this is the molecular chaos assumption). With
these assumptions, in the so-called Grad-Boltzmann limit, one gets the Boltzmann equation.
However, for the Boltzmann equation, time has an arrow and the dynamics is irreversible,
unlike the Hamiltonian equation that governs fN . This inconsistency reflects the fact that
atomistic dynamics is not the only mechanism that should be taken into account, and in fact
all the other influences are subtly hidden in the assumptions made in the derivation, the details
of which is another topic that we will not discuss here.
The Boltzmann equation reads:
∂tf + v · ∇xf =1
εQ(f, f), t ≥ 0, (x, v) ∈ Rd × Rd. (4.1.1)
d is the dimension. Here ∂tf + v · ∇xf part is to formulate the free transport: particles move
in a straight line with velocity v. The Q is called the collision term, and in general has the
following form:
Q(f, f) = Q+ − fQ− =
∫Sd−1
∫Rd
(f ′f ′∗ − ff∗)B(|v − v∗|, ω)dv∗dω. (4.1.2)
This term means that two particles with pre-collisional velocities, one with v and another with
v∗, bump into each other, and after colliding, they change their velocities to post-collisional
velocities: v′ and v′∗. If the collision is elastic, momentum and energy are conserved, i.e.
v + v∗ = v′ + v′∗, |v|2 + |v∗|2 = |v′|2 + |v′∗|2,
93
Figure 4.1.1: Velocity change after the collision.
and this leads to:
v′ = v − 1
2(g − |g|ω), v′∗ = v∗ +
1
2(g − |g|ω). (4.1.3)
where ω is a unit vector on Sd−1, the unit sphere defined in Rd space. See figure 4.1.1. We
use the shorthands f ′ = f(t, x, v′) and f ′∗ = f(t, x, v′∗). Q+ and fQ− parts are called gaining
and losing part respectively, reflecting how many more or fewer particles having the velocity
v. B is called collisional cross-section, or collision kernel. It is always positive, and is used to
measure the repartition of post-collisional velocities given the pre-collisional velocities. It varies
according to the nature of particles considered. For hard sphere particles, B(z, ω) ∼ z. For
particles governed by power law, B(z, ω) = b(ω)|z|γ , and specifically when γ = 0, the particles
are called the Maxwell molecule. The Maxwell molecule is not physical but is used very often
for the purpose of obtaining explicit calculation in agreement with physical observables. In
physics, numerous examples have grazing collision [50, 47], and possess non-integrable b(ω),
so mathematically we have to go through “cut-off” process by replacing it with some bounds
that are integrable [87].
There are several assumptions we made: firstly, we assume particles change direction only
through colliding with other particles. Since collision is short range, the interaction is very
sensitive to the exact value of both position and velocity. If long range potential is considered,
the equation would undergo dramatic change, for example one gets Vlasov equation in Coulomb
94
case where collision term is replaced by effective field [26]. Secondly, we have only considered
binary collisions: in rarefied gas, the collisions that involve more than three particles are so
rare and are simply ignored. Finally, we assume the collision is elastic, and this does not have
to be the real case: energy decay sometimes takes place for granular gases when collisions are
inelastic [187, 18, 32, 71].
4.1.1 Properties of the collision term
In this subsection we state some important properties of the collision term [33]. All the
derivations are purely formal.
Conservation law
Cross-section may vary, but the first d+ 2 moments of the collision term are always zero.
In fact, given arbitrary smooth function φ, we have:
∫Q(f, f)φ(v)dv =
∫B(f ′f ′∗ − ff∗)φdvdv∗dω
=1
2
∫B(f ′f ′∗ − ff∗)(φ+ φ∗)dvdv∗dω
=1
4
∫B(f ′f ′∗ − ff∗)(φ+ φ∗ − φ′ − φ′∗)dvdv∗dω
In the derivation, we used that the cross-section B is invariant to |v− v′|. For elastic collision,
immediately, one has that the quadratic functions in v are collisional invariants: let φ =(1, v, 1
2 |v|2)T
, the first d+ 2 moments are zero:
< Q > =
∫Q(f)dv = 0,
< vQ > =
∫vQ(f)dv = 0,
<1
2v2Q > =
∫1
2|v|2Q(f)dv = 0. (4.1.4)
95
Based on these formulas, we take moments of the Boltzmann equation and obtain conser-
vation laws. To do that, we firstly define the following macroscopic quantities:
ρ =
∫fdv, ρu =
∫vfdv,
E =1
2ρu2 + ρe =
1
2
∫|v|2fdv, e =
3
2T =
1
2ρ
∫f |v − u|2dv, (4.1.5)
S =
∫(v − u)⊗ (v − u)fdv, q =
1
2
∫(v − u)|v − u|2fdv.
ρ is called local density, T is local temperature and u is local bulk velocity. All these quantities
are “local” in space. With these, we have:
∂tρ+∇x · (ρu) =< Q >= 0,
∂t(ρu) +∇x · (S + ρu2) =1
ε< vQ >= 0, (4.1.6)
∂tE +∇x · (Eu+ Su+ q) =1
ε<
1
2|v|2Q >= 0.
H theorem
The conservation law does not reflect the built in time-irreversibility of the Boltzmann
equation. The irreversibility is seen through the so called “Boltzmann’s H-theorem”, which
basically says that entropy of a system keep decaying in time. We multiply Q with log f .
Simple algebra gives:
∫Q(f, f) log fdv =
1
4
∫B(f ′f ′∗ − ff∗) log
ff∗f ′f ′∗
dvdv∗dω ≤ 0
Note again the discussion on integrability and such are all waived, and the computation is
purely formal.
From here we have two conclusions.
96
• Entropy decay. Define entropy as∫f log fdv, then we have:
∂t
∫f log fdv +∇x ·
∫vf log fdv =
∫Q(f, f) log fdv ≤ 0 (4.1.7)
• Entropy achieves its minimum when the system gets to the equilibrium state. Mathemat-
ically we set∫Q(f, f) log fdv = 0, and the solution to this is that log f is a quadratic
function, i.e. f is a Gaussian distribution. We call the Gaussian function that are
equipped with proper macroscopic quantities Maxwellian (macroscopic quantities are
computed from f and thus f appears as the subindex):
Mf (t, x, v) = ρ(t, x)
(1
2πT (t, x)
)d/2exp
(−(v − u(t, x))2
2T (t, x)
). (4.1.8)
4.1.2 Non-dimensionalization and rescaling process
In different regimes, one observes the system with different time and length scales and
acquires different behavior of the solution. We firstly perform the non-dimensionalization.
Denote l0 and t0 as length and time scale, and a reference temperature is marked as T0. The
thermal speed is defined out of it:
c ∼√kbT0
m(4.1.9)
with kb being the Boltzmann constant and m is the molecular mass. Rescale the length and
time as following:
t =t
t0, x =
x
l0, and v =
v
c. (4.1.10)
97
Another time scale associated with the intensity of the collision cross-section is called mean
free time τ , a parameter used to reflect the time a particle undergoes between two collisions.
∫Mf (v)Mf (v∗)Bdσdv∗dv =
N
l30τ, (4.1.11)
where Mf is the Maxwellian defined in (4.1.8). Evidently, bigger B means that the system has
more frequent collisions, and thus the mean free time is shorter. Mean free path is
λ = cτ, (4.1.12)
reflecting the distance a particle travels between two collisions. Finally we re-normalize the
number density and the collision kernel as well:
f(t, x, v) =l30c
3
Nf(t, x, v), B(v − v∗, σ) = ρτB(v − v∗, σ). (4.1.13)
The equation thus becomes (dropping all the tildes):
l0ct0
∂tf + v · ∇xf =l0λQ(f, f). (4.1.14)
Define the Knudsen number as the mean free path over typical domain length λl0
, denoted by
Kn. It is divided from the collision term Q. The ratio in front of the ∂tf term is called the
Strouhal number, denoted by St. We tune these two parameters and could observe various
phenomena in different regime. In general, as Kn → 0, one gets the Euler equations in the
first asymptotic limit, and the Navier-Stokes equation in the second. If St is similarly small,
the macroscopic equation are incompressible; and if it is O(1), one gets compressible limit.
98
4.2 Asymptotic limit
We only derive the compressible Euler limit in this thesis, i.e. Knudsen number is small,
replaced by ε, while the Strouhal number is of O(1):
∂tf + v · ∇xf =1
εQ(f, f). (4.2.1)
While deriving for asymptotic limit of the Boltzmann equation, different ansatz would give
different hierarchies of PDEs. In history, there have been two dominating approaches: one is
called Hilbert expansion, that dates back to Hilbert’s original work in [101], in which he sought
for solutions of formal power series in ε:
f(t, x, v) =∑k
εkfk(t, x, v). (4.2.2)
Another approach is called Chapman-Enskog expansion developed independently by Chap-
mann and Enskog [37]. Their idea is to expand the solution close to the local Maxwellian:
f(t, x, v) = Mf
(1 +
∑k>0
εkgk(t, x, v)
). (4.2.3)
Here Mf is the local Maxwellian function given in (4.1.8), and g is regarded as perturbation
that should vanish at infinite time. These two approaches give different expansion processes,
but in the leading order of ε, both of them imply that the distribution function should be
a Gaussian function, and in particular, Chapmann-Enskog expansion assign this Gaussian
function the local Maxwellian:
f0 = Mf = ρ(t, x)
(1
2πT (t, x)
)d/2exp
(−(v − u(t, x))2
2T (t, x)
),
99
with ρ, T u defined as in (4.1.5). The idea of Chapmann-Enskog expansion is that since Mf
and f has the same moments, and thus one could plug it back into (4.1.6) and close the system.
In this leading order, the compressible Euler equation is obtained:
∂tρ+∇ · (ρu) = 0,
∂t(ρu) +∇ · (ρu⊗ u+ ρT I) = 0, (4.2.4)
∂tE +∇ · ((E + ρT )u) = 0.
This finishes the leading order asymptotic expansion. In the next order we obtains compressible
Navier-Stokes limit, by plugging f = Mf (1 + εg1) back into the original Boltzmann equation,
and solving for the kernel of linearized Boltzmann collision term. We omit the process here.
Remark 4.2.1. A side note on rigorous proofs on the Boltzmann equation. In general there
are three ways in connecting the two set of equations. One approach is based on truncated
asymptotic expansions. Rigorous justification is done by Caflisch [30] for compressible Euler
equation, up to the appearance of the singular time for the Euler equation. Similar result for
incompressible Navier-Stokes equations is obtained by De Masi, Esposito and Lebowitz [44],
where they also require smoothness of the solutions to the macroscopic equations. Another
approach is under the perturbation framework: one expands the distribution function around a
renormalized target Maxwellian (in infinite time), and analyze the spectrum of the linearized
Boltzmann equation. The spectrum was firstly explored by Ellis and Pinskii [59]. The idea
was later used by Nishida for compressible Euler equations [159], and by Bardos and Ukai for
incompressible Navier-Stokes equations [8]. As in the first approach, the results obtained rely
on the existence of classical solutions to the macroscopic equations, and thus either becomes
invalid in finite time or requires smallness on the initial data. However, physically, one should
expect that the existence and smoothness of the Boltzmann equation should not rely on that
100
of the macroscopic equations. In fact, the global existence is done in a series of work by
Lions, initiated in [54] and summarized in [146]. To get rid of these non-physical assumptions,
Bardos, Golse and Levermore proposed the third approach [6, 5], which is to link the two sets
of equations in the weak sense, and they sought for possibility to obtain Leray [140] solutions
to incompressible Navier-Stokes equation. The program was complete by Golse and Saint-
Raymond in [78, 77]. None of these theoretical works will be discussed in this thesis, but they
are milestones in the development of the whole theory.
101
Chapter 5
Asymptotic preserving numerical
methods
In this chapter we investigate the asymptotic preserving numerical methods. Before going
into details of two methods in the following two sections, we firstly briefly introduce this
concept.
Usually multi-scale problems contain some parameters that could vary across several scales.
As the parameters are manually sent either big or small, one carries asymptotic expansions
on these parameters, yielding asymptotic limits that are described by averaged or effective
equations. In our case, the Euler equations are the asymptotic limit of the Boltzmann equation
as the Knudsen number goes to zero.
Numerically, in these asymptotic regimes, the original equation could be difficult to com-
pute: the small scales need to be resolved in discretization, generating tremendous compu-
tational cost. So whenever possible, we tend to simply solve the asymptotic or macroscopic
equation as it is more efficient. However, in practise, usually the macroscopic model breaks
down in part of the domain, and two or more regimes co-exist. When this happens, a natu-
ral idea is to do decompose the domain into parts and compute different sets of equations in
different regions. Under this framework, one needs to know specifically where to set up the
boundary and how to couple the two systems, but neither is easy. Some kinetic equations can
be computed in this line of thinking including [23, 45, 46, 76, 128, 138], but the idea does not
102
work well for the Boltzmann equation. In fact, understanding the structure of the boundary
layer is a challenging problem unsolved for decades by itself [14, 7, 4].
Another approach is to develop asymptotic-preserving (AP) schemes, an approach that we
are pursuing in this thesis. It firstly appeared for neutron transport equation in the diffusive
regime [134, 135], and was widely used later on in many other models. The idea of this concept
is to develop cheap numerical methods only for the microscopic equation, and the methods
should be robust in all regimes, so they automatically capture the asymptotic limit in the
stiff regime, with fixed discretization. More accurate error analysis could be found in [111].
Specifically, in our case, for the Boltzmann equation, as summarized by Jin [110], a scheme is
AP if:
• it preserves the discrete analogy of the Chapman-Enskog expansion, namely, it is a
suitable scheme for the kinetic equation, yet, when holding the mesh size and time step
fixed and letting the Knudsen number go to zero, the scheme becomes a suitable scheme
for the limiting fluid dynamic Euler equations;
• implicit collision terms can be implemented efficiently.
Typically, to relax the time step from the control of the Knudsen number, one needs to use an
L-stable numerical scheme, and thus the collision term should be computed implicitly. How-
ever, as seen above, the Boltzmann collision term is non-local and nonlinear, and numerically
inverting it is impossible. All of these conflicts are making the designing of AP schemes for
the Boltzmann equation difficult.
There are several variations of the AP property, with some of them easier to achieve than
others. They are weakly-AP, relaxed-AP, and strongly-AP, defined as follows (see [65]):
• weakly-AP. If the data are within O(ε) of the local equilibrium initially, they remain so
for all future time steps;
103
• relaxed-AP. For non-equilibrium initial data, the solution will be projected to the local
equilibrium beyond an initial layer (after several time steps).
• strongly-AP. For non-equilibrium initial data, the solution will be projected to the local
equilibrium immediately in one time step.
In general, the strongly-AP property is preferred, and was the designing principle for most
classical AP schemes [109, 29]. The relaxed-AP is a concept introduced recently in [65], which
was shown numerically to be sufficient to capture the hydrodynamic limit when the Knudsen
number goes to zero. The weakly-AP is even weaker: it requires the solutions do not move
away from the equilibrium if they are close to it initially. It serves as a necessary condition for
the AP property.
We mention here several AP schemes recently developed. One approach was to use the
micro-macro decomposition method [139] (see its multi-species extension in [119]), but the
issue of designing an efficient implicit collision term, which is necessary for numerical stability
independent of the Knudsen number, is still unsolved. An earlier approach introduced by
Gabetta, Pareschi and Toscani uses the truncated Wild Sum for uniform numerical stability
of the collision term [70]. In the following sections, we introduce in detail two methods.
The first one is the BGK penalization introduced in [65] and its stronger version can be
found in [199]. The idea was later on extended to hyperbolic systems [68], Fokker-Planck-
Landau equation [123], quantum Boltzmann equation [64], quantum Fokker-Planck-Landau
equation [104] and multi-species system [112]. Higher order accuracy in time is achieved with
IMEX in [53]. The second one is the Exponential Runge-Kutta method, which was developed
in [52] for homogeneous equation, and later on generalized by Pareschi and the author for
space inhomogeneous case [143].
104
5.1 BKG-penalization method
This approach was firstly developed in [65]. The idea is to use a BGK operator β(Mf −
f) [15] to approximate the collision term Q(f, f) and the difficulty in inverting Q is transfered
to implicitly computing the BGK operator. The detail is the following: we split the collision
operator into the BGK part and the rest. Since the BGK term is a good approximation to
the collision term, after the subtraction, the remainder is less stiff, and is treated explicitly,
and the BGK term is stiff, and should be computed implicitly. However, although it is also
a nonlinear term on f , and is not invertible either, it is discovered in [42, 167] that this term
could be computed in an explicit manner. We firstly rewrite the equation in the following:
∂tf + v · ∇xf =Q(f)− Pb(f)
ε+Pb(f)
ε. (5.1.1)
with Pb = β(Mf − f) is a BGK term standing for penalization, and β is constant to be
determined. Pb(f) is expected to cancel the leading stiff order in Q, the first term on the
right is less stiff, and one could use explicit methods to handle it without worrying about the
stability requirements on time discretization. The second term however, needs to be treated
implicitly, for example:
f l+1 − f l∆t
+ v · ∇xf l =Q(f l)− Pb(f l)
ε+Pb(f
l+1)
ε. (5.1.2)
The superscript l stands for the time step. We also use shorthands Ql , Q(f l), P lb , Pb(fl)
for convenience. With simple algebra, one gets:
f l+1 =εf l + ∆t(Ql − βl(M l
f − f l))− εhv · ∇xf l + βl+1hM l+1f
ε+ βl+1h. (5.1.3)
105
h is time discretization. The computation of v ·∇xf is very mature, and the reader could refer
to [141] for varies numerical shock-capturing highly accurate methods. The computation of
Q(f) is also a topic that have been studied through years [156, 165, 67]. We mainly focus on
computing Mf and selection of β in the following.
• The computation of M l+1f : One only needs the evaluation of the macroscopic quan-
tities for the next time step in order to compute Mf . To do that, we take the first d+ 2
moments numerically of the numerical scheme (5.1.2). The moments on the right hand
side disappear and one gets:
ρl+1 = ρl − h∫v · ∇xf l dv,
(ρu)l+1 = (ρu)l − h∫v ⊗ v∇xf l dv,
El+1 = El − h∫
1
2|v|2v · ∇xf l dv,
and the temperature T l+1 is simply: T l+1 = 2El+1−(ρu2)l+1
dnl+1 . With these, Mf is defined
by (4.1.8).
• The selection of β: The selection of β could be tricky. Generically, it plays a role of
Frechet derivative:
Q(f) ∼ Q(Mf ) +∇Q(M)[M − f ]. (5.1.4)
Considering Q(Mf ) = 0, β should be chosen to simulate ∇Q(f). In fact, often in time,
one chooses βl = sup∣∣∣ Q(f)f−Mf
∣∣∣, or βl = sup∣∣∣Ql−Ql−1
f l−f l−1
∣∣∣. Later on in [199], the authors
discovered that we could tune this parameter to ensure the positivity of the distribution
106
function. To be more clear, if β satisfies the following:
βl = sup Q−(f l)
βl+1 = βl(1 + hκl), with κl = max
sup
M lf−M
l+1f
hM l+1f
, 0
.
then the distribution function in (5.1.3) is positive.
This method was proved to be relaxed-AP in [65], and the strong-AP version is obtained
in [199]. The proof will be omitted from here.
5.2 Exponential Runge-Kutta method
This is another AP method proposed in [52] that essentially could achieve arbitrary high
order of accuracy. It was designed for the spatial homogeneous Boltzmann equation, namely
the transport term v · ∇xf is not included. For space inhomogeneous case, one has to use
time-splitting, for example, the Strang splitting, which is 2nd order in time, therefore the high
accuracy obtained in the collision step is lost. This drawback was later on fixed in [143] where
the authors directly manipulate the entire space inhomogeneous Boltzmann equation without
splitting. The immediate consequence is the arbitrarily high order of accuracy.
The idea is to reformulate the equation in a way so that the explicit methods, when applied
upon them, automatically pushed the distribution function to the right Maxwellian. Some new
parameters appear in the reformulation, and numerically one needs to select values for them
in a smart way. In fact, up to now two choices are found, both of which are proved to lead
to AP. In the following subsections, we describe the reformulation of the equation and present
the two numerical methods in 5.2.1, followed with the numerical properties in subsection 5.2.2.
The numerical examples will be given in subsection 5.2.3.
107
5.2.1 Numerical methods
In this subsection we firstly reformulate the Boltzmann equation so that the Exponential
RK method could be carried out, and then introduce the two numerical schemes.
Reformulation of the problem and notations
We need to formulate the equation in a way such that generic explicit methods, as long
as consistent, could also preserve the right asymptotic limit. Inspired by [52], we have the
following:
∂t
[(f − M)eµt/ε
]= ∂t(f − M)eµt/ε + (f − M)µε e
µt/ε (5.2.1)
=[
1ε (Q+ µf − µM)− ∂tM − v · ∇xf
]eµt/ε (5.2.2)
. =[
1ε (P − µM)− ∂tM − v · ∇xf
]eµt/ε.
In the derivation, µ is a constant, and M is an arbitrary function, called equilibrium function,
the scheme for whom will be carefully designed later. Note that M is not the local Maxwellian
Mf . P is usually regarded as the gain part of the collision, defined as:
P = Q+ µf. (5.2.3)
The equation obtained is always equivalent to the original Boltzmann equation, regardless of
the form of M and µ, but numerically, as we will see later, to capture the correct asymptotic
limit, they are selected under restrictive criteria. In fact, if one applies the simple forward
Euler scheme onto it, clearly one gets f going to M exponentially fast. So to realize asymptotic
preserving properties, M has to be chosen close enough to the real limit, the local Maxwellian
Mf . To do that, we have two methods, and depending on whether or not M change with
respect to time in each time step, we have ExpRK-F (standing for fixed M) and ExpRK-V
108
(varying M) respectively.
Exponential RK schemes with fixed equilibrium function
In the equation (5.2.1), it is easy to see that the trickiest term is the time evolution of the
equilibrium function ∂tM . So in our first try, we explore possibility to get rid of this term, i.e.
we look for an a-priori constant function so that ∂tM is manually erased from the equation in
each time step. If succeed, (5.2.1) is modified accordingly:
∂t
[(f − M)eµt/ε
]=
[1
ε(P − µM)− v · ∇xf
]eµt/ε. (5.2.4)
As argued above, M need to be chosen close enough to the Maxwellian Mf . However, as one
could expect, this asymptotic limit itself is a function of time and ∂tMf 6= 0. So we are going
to numerically evolve Mf firstly, and set M as the updated Mf for each time step. The general
explicit Runge-Kutta framework to numerically solve (5.2.4) is:
Step i: (f (i) − M)eciλ = (f l − M) +
i−1∑j=1
aijh
ε
[P (j) − µM − εv · ∇xf (j)
]ecjλ
Final Step: (f l+1 − M)eλ = (f l − M) +
ν∑i=1
bih
ε
[P (i) − µM − εv · ∇xf (i)
]eciλ
,
(5.2.5)
109
where M is defined by the macroscopic quantities (ρl+1, ul+1, El+1) computed from:
Step i :
ρ
ρu
E
(i)
=
ρ
ρu
E
l
− hi−1∑j=1
aij∇x ·
ρu
ρu⊗ u+ ρT
(E + ρT )u
(j)
Final Step:
ρ
ρu
E
l+1
=
ρ
ρu
E
l
− hν∑i=1
bi∇x ·
ρu
ρu⊗ u+ ρT
(E + ρT )u
(i) . (5.2.6)
We explain the notations above, ν is the number of stages of the Runge-Kutta method, with
aij and bi as its coefficients. ci =∑
j aij . Superscript n denote the l-th time step, and (i)
denote the sub-stage i in l-th time step. h is the time discretization, and P (j) = P (f (j)) for
simplicity.
Remark 5.2.1.
• Note that this method gives us a simple way to couple macro-solver with micro-solver.
When ε is considerably big, the accuracy of the method is controlled by the micro-solver.
And as ε vanishes, the method pushes f going to M , which is defined by macroscopic
quantities computed through the Euler equation while the order of accuracy is given by
the macro-solver.
• In principle it is possible to adopt other strategies to compute a more accurate time
independent equilibrium function in intermediate regions. For example one can use the
ES-BGK Maxwellian [66] at time n + 1, or one can use the Navier-Stokes equation as
the macro-counterpart. Here however we do not explore further in these directions.
In each time step, we update macroscopic quantities (5.2.6) first, and use them to define
110
M . Plug it back into (5.2.5) and update the distribution function f .
Exponential Runge-Kutta schemes with time varying equilibrium function
In this approach we keep the ∂tM term and try to find an efficient numerical method to
compute it. In fact, we consider the most natural choice of equilibrium function, namely the
local Maxwellian M = Mf . Therefore, M is a Gaussian profile that has the same first d + 2
moments with f , and the evolution of the moments are governed by:
∂t
∫φfdv +
∫φv · ∇xfdv = 0, (5.2.7)
with φ =[1, v, v
2
2
]T. The RK methods are:
Step i:
(f (i) −M (i))eciλ = (f l −M l)
+i−1∑j=1
aijh
ε
[P (j) − µM (j) − εv · ∇xf (j) − ε∂tM (j)
]ecjλ,
∫φf (i)dv =
∫φf ldv +
i−1∑j=1
aij
(−h∫φv · ∇xf (j)dv
);
(5.2.8a)
Final Step:
(f l+1 −M l+1)eλ = (f l −M l)
+
ν∑i=1
bih
ε
[P (i) − µM (i) − εv · ∇xf (i) − ε∂tM (i)
]eciλ,
∫φf l+1dv =
∫φf ldv +
ν∑i=1
bi
(−h∫φv · ∇xf (i)dv
).
(5.2.8b)
Note that the coupling between the two equation is embedded: M (j) are computed through∫φf (j)dv, and will be used to update f (i) for i > j. In each sub-stage i, one needs to evaluate
111
the following: f (j), M (j), ∂tM(j), P (j), v ·∇xf (j) for all j < i, and M (i) that is evaluated at the
current time sub-stage. f (j) and M (j) are obtained from previous time sub-stages, and P (j)
could be computed through spectral method developed in [165], for the N logN improvement,
see [156]. v ·∇xf (j) is computed using standard ENO or WENO method to match up the time
accuracy designed. We focus on computation of M (i) and the tricky term ∂tM :
Computation of M (i) :
solve the second equation of (5.2.8a), to get evaluation of macroscopic quantities at
tl + cih. Then the Maxwellian M (i) is given by (4.1.8).
Computation of ∂tM(j) :
Write ∂tM as
∂tM = ∂ρM∂tρ+∇uM · ∂tu+ ∂TM∂tT, (5.2.9)
and ∂tρ, ∂tu and ∂tT can be computed from taking moments of the original equation
∂t
ρ
ρu
dρT2 + 1
2ρu2
= ∂t
∫
1
v
v2
2
Mdv = ∂t
∫
1
v
v2
2
fdv
(5.2.10)
= −∫
1
v
v2
2
v · ∇xfdv.
To be specific, with data at sub-stage (j) in d-dimensional space, one has
∂tM(j) = ∂ρM
(j)∂tρ(j) +∇uM (j) · ∂tu(j) + ∂TM
(j)∂tT(j), (5.2.11)
112
with
∂ρM(j) =
M (j)
ρ(j), ∂uM
(j) = M (j) v − u(j)
T (j),
∂TM(j) = M (j)
[(v − u(j))2
2(T (j))2− d
2T (j)
],
and all the macroscopic quantities associated with f (j), namely ρ(j), u(j) and T (j) are
given by:
∂tρ(j) = −
∫v · ∇xf (j)dv,
∂tu(j) =
1
ρ(j)
(u(j)
∫v · ∇xf (j)dv −
∫v ⊗ v · ∇xf (j)dv
),
∂tT(j) =
1
dρ(j)
(−2E(j)
ρ(j)∂tρ
(j) − 2ρ(j)u(j)∂tu(j) −
∫v2v · ∇xf (j)dv
).
5.2.2 Properties of ExpRK schemes
The numerical schemes we designed possess some very nice features. We are especially
interested in positivity and asymptotic preserving.
Positivity and monotonicity properties
The distribution function describes the number of particles at certain phase point, and
should keep being positive along the evolution. However, it is hard to maintain positivity
numerically, especially when the schemes used are high order. Luckily, our ExpRK-F method,
under some very mild assumption, shows positivity. To that end, we use a very powerful tool
discovered in [191] and called Shu-Osher representation, and the analysis follows [83, 84]. We
review Shu-Osher representation first before presenting the proof.
For a large set of ODEs:
∂ty = F (t, y), (5.2.12)
113
The general Runge-Kutta writes as:
Step i: yl,(i) = yl + h
i−1∑j=1
aijF (tl + cjh, yl,(j))
Final step: yl+1 = yl + hν∑i
biF (tl + cih, yl,(i))
. (5.2.13)
We intend to represent it as:
Step i: y(i) =
i−1∑j=1
[αijy
(j) + hβijF (tl + cjh, y(j))],
Final step: yl+1 =
ν∑j=1
[αν+1jy
(j) + hβν+1jF (tl + cjh, y(j))].
(5.2.14)
Note that there are more parameters one could adjust in (5.2.14) than in (5.2.13), and to
make these two formulas equivalent, the selection of coefficients in (5.2.14) is not unique. For
consistency, we require∑i−1
j=1 αij = 1, and for β, we use the most natural choice:
βij = αij (ci − cj) . (5.2.15)
Through simple algebra, one could prove that this choice is equivalent to the original one
proposed in [191] where βij = aij−∑αikakj . In fact, assume one has y(j) = yl+h
∑j−1k=1 ajkF
(k),
∀ j < i, where F (k) is a shorthand for F (tl + ckh, y(k)), then, our choice gives:
y(i) =i−1∑j=1
[αijy
(j) + αij(ci − cj)hF (j)]
=∑j<i
αijyl + h
∑k<j
ajkF(k)
+ αij(ci − cj)hF (j)
= yl + h
∑j<i
i−1∑k=j+1
αikakj + αij(ci − cj)
F (j) (5.2.16)
114
This clearly requires aij = αij(ci − cj) +∑αikakj , which is exactly the classical Shu-Osher
representation: aij = βij +∑αikakj .
With this, our numerical scheme could be transformed into:
f (i) − M)eciλ =
∑j
ecjλαij(f
(j) − M) + βijh
ε
[P (j) − µM − εv · ∇xf (j)
]
(f l+1 −M)eλ =∑j
ecjλαν+1j(f
j − M) + βν+1jh
ε
[P (j) − µM − εv · ∇xf (j)
] .
We claim under very mild assumption, the ExpRK-F method gives positive distribution
function. The assumption we are going to make is:
Assumption 5.2.1. For a given f > 0 there exists h∗ > 0 such that
f − h v · ∇xf > 0, ∀ 0 < h ≤ h∗.
Note that this assumption is very mild, and is the minimal requirement on f in order to
obtain a non negative scheme. The assumption is naturally valid since f − v · ∇xf is simply
the forward Euler scheme for a transport equation with constant coefficient (v in this case).
Next we can state:
Theorem 5.2.1. Consider an ExpRK-F method defined by (5.2.5), and βij ≥ 0 in (5.2.15).
There exist h∗ > 0 and µ∗ > 0 such that f l+1 ≥ 0 provided that f l ≥ 0, µ ≥ µ∗ and 0 < h ≤ h∗.
Proof. We prove it for sub-stage i, and the analysis for the final step could be carried out in
115
the same manner. Simple algebra gives, for i = 1, · · · , ν, j < i
f (i) =M
1−∑j
e(cj−ci)λ (αij + λβij)
+
i−1∑j=1
λβije(cj−ci)λP
(j)
ε
+i−1∑j=1
αije(cj−ci)λ
(f (j) − hβij
αijv · ∇xf (j)
). (5.2.17)
To prove positivity, it is enough to show that this is a convex combination, and each element
is positive. To show each element is positive, i.e.
M > 0, P (j) > 0, f (j) − hβijαij
v · ∇xf (j) > 0. (5.2.18)
Positivity of M is obvious, and P (j) is positive if one has big enough µ:
µ ≥ µ∗ = sup |Q−| ⇒ P = Q+ µf = Q+ − fQ− + µf > 0.
To handle the positivity of the transport term, we adopt Assumption 5.2.1. The term is
positive as long as h∗ is small enough:
0 < h ≤ h∗ = minij
(αijβij
h∗).
We also check the convexity of (5.2.17), it should be proved that
∑j
e(cj−ci)λ (αij + λβij) ≤ 1. (5.2.19)
In fact, it is easy to show that the quantity on the left hand side is a decreasing function for
λ ≥ 0, that gets its maximum value∑
j αij = 1 at λ = 0. One could simply take the derivative
116
with respect to λ and will find it is negative:
d
dλ
∑j
e−∆ijλ(αij + λβij)
=∑j
e−∆ijλ (−∆ij(αij + λβij) + βij) (5.2.20)
=∑j
e−∆ijλ (−βij∆ijλ+ βij − αij∆ij) < 0 (5.2.21)
Here ∆ij = ci − cj . In the last step, βij = αij(ci − cj) is used. This finishes the proof on
convexity and we conclude that the scheme gives positive distribution function in each time
step.
Since the proof above is based on a convexity argument, we also have monotonicity of the
numerical solution or SSP property. Thus the building block of our exponential schemes is
naturally given by the optimal SSP schemes which minimize the stability restriction on the
time stepping. We refer to [84] for a review on SSP Runge-Kutta schemes.
Remark 5.2.2.
• Note that the proof above does not rely on the value λ take, i.e. the scheme is positive
uniformly in ε.
• Optimal second and third order SSP explicit Runge-Kutta methods such that βij ≥ 0 have
been developed in the literature. However the classical third order SSP method by Shu and
Osher [191] does not satisfy cj ≤ ci for j < i. Note that standard second order midpoint
and third order Heun methods satisfy the assumptions in the Theorem 1(see Table 1.1
page 135 in [98]).
In [83] it was proved that all four stage, fourth order RK methods with positive CFL
coefficient h∗ must have at least one negative βij. The most popular fourth order method
117
using five stage with nonnegative βij has been developed in [176]. In [176] the authors
also proved that any method of order greater then four will have negative βij.
• Positivity of ExpRK-V schemes is much more difficult to achieve because of the involve-
ment of the ∂tM term. However, we can prove
– ρ is positive;
– the negative part of T is O(hε).
We leave both proofs to the appendix B.1.
Contraction and Asymptotic Preservation
In this section, we show ExpRK-F and ExpRK-V are asymptotic preserving numerical
method. The whole idea relies on the contraction property of the collision operator P [52].
Formally, if one check the formulas (5.2.5) and (5.2.8b), as ε → 0, λ → ∞, and the
exponential functions as a penalty that pushes f approaching to M . The two methods above
assign M with macroscopic quantities that numerically solve the Euler equation, and thus AP
follows directly. The proof in this section is carried out in a rigorous manner: we carefully track
in each time step the difference between the distribution function and the correct Maxwellian
function and provide accurate estimation on the convergence rate. Especially our proof shows
that the argument could also handle the tough case, namely, cν = 1. When this happens, the
formal argument does not carry through.
The proof relies on the following assumption:
Assumption 5.2.2. There is a constant C big enough, such that |P (f, f)− P (g, g)| < C |f − g|
where |·| denotes a proper metric.
This assumption is on the contraction of the operator P , and is generally true for most
physically meaningful metric. In appendix B.2, we show the validity of the proof with the
118
metric d2 in space Ps(Rd) (also see [183]).
Under this assumption, considering P (M,M) = Q(M,M) + µM = µM , one has
|P (f, f)− µM | < C |f −M | . (5.2.22)
The proofs for both approaches: ExpRK-F and ExpRK-V are exactly the same, and we only
show it for ExpRK-F for simplicity, with any given explicit Runge-Kutta scheme.
For AP property, one needs to show that as ε→ 0, the scheme gives correct Euler limit. To
do this, basically one needs to prove that f goes to the Maxwellian function whose macroscopic
quantities solve the Euler equation (4.2.4).
Let us define
di =∣∣∣f (i) − M
∣∣∣ , Di =∣∣∣v · ∇xf (i)
∣∣∣ , d0 =∣∣∣f l − M ∣∣∣ , ~e = [1, 1, · · · , 1]T ,
~d = [d1, d2, · · · , dν ] , ~D = [D1, D2, · · · , Dν ]T . (5.2.23)
Moreover A is a lower-triangular matrix and E is a diagonal matrix given by
Aij =λ
µaije
(cj−ci)λ, E = diage−c1λ, e−c2λ, · · · , e−cνλ.
Lemma 5.2.1. Based on the definitions above, for ExpRK-F one has
~d ≤ d0 (I− CA)−1 · E · ~e+ ε (I− CA)−1 · A · ~D.
Proof. It is just direct derivation from (5.2.5):
(f (i) − M)eciλ = (fn − M) +
i−1∑j=1
aijλ
µecjλ(P (j) − µM − εv · ∇xf (j)). (5.2.24)
119
We take the norm of the equation and use the triangle inequality. With the contraction
assumption, one gets:
∣∣∣f (i) − M∣∣∣ ≤ ∣∣∣f l − M ∣∣∣ e−ciλ +
∑j
aijλ
µe(cj−ci)λ
(C∣∣∣f (j) − M
∣∣∣+ ε∣∣∣v · ∇xf (j)
∣∣∣) . (5.2.25)
A condense matrix form is:
d1
d2
...
dν
≤ E
d0
d0
...
d0
+ CA
d1
d2
...
dν
+ εA
D1
D2
...
Dν
.
So
~d ≤ d0E · ~e+ CA · ~d+ εA · ~D (5.2.26)
~d ≤ d0 (I− CA)−1 · E · ~e+ ε (I− CA)−1 · A · ~D (5.2.27)
which completes the proof.
Lemma 5.2.2. ∣∣∣f l+1 − M∣∣∣ ≤ ∣∣∣f l − M ∣∣∣R1(λ) + ~R2 · ~D, (5.2.28)
with R1 and R2 defined as:
R1(λ) = e−λ(
1 +Cλ
µ~b · E−1 (I− CA)−1 E · ~e
)(5.2.29)
~R2(λ) =ελ
µe−λ~b · E−1 · (I− CA)−1 · (I + CA) (5.2.30)
120
Proof. It is just a simple derivation. Define
ki =h
ε(P (i) − µM − εv · ∇xf (i))eciλ. (5.2.31)
Apparently, the previous lemma leads to
~|k| ≤ λ
µE−1 ·
(C~d+ ε ~D
). (5.2.32)
Back to (5.2.5), one has
(f l+1 − M
)=(f l − M
)e−λ +
ν∑s=1
bikie−λ, (5.2.33)
which implies
∣∣∣f l+1 − M∣∣∣ ≤d0e
−λ +λ
µe−λ~bT · E−1 ·
(C~d+ ε ~D
)(5.2.34a)
≤e−λ(d0 +
λ
µ~b · E−1 ·
(C (I− CA)−1 ·
(d0E · ~e+ εA · ~D
)+ ε ~D
))(5.2.34b)
≤d0e−λ(
1 +Cλ
µ~b · E−1 · (I− CA)−1 · E · ~e
)(5.2.34c)
+ελ
µe−λ~b · E−1 · (I− CA)−1 · (I + CA) · ~D. (5.2.34d)
Here ~b = [b1, b2, · · · , bν ] is a row vector. The result (5.2.27) is also used. We conclude with the
definition of R1/2. ∣∣∣f l+1 − M∣∣∣ ≤ ∣∣∣f l − M ∣∣∣R1(λ) + ~R2(λ) · ~D.
The two lemmas are pure algebra computation, but they provide the exact formula on
how distribution function convergence towards the Maxwellian. Apparently the smaller R1 is,
121
the faster the function converges, and R2 represents the drift from the transportation, and is
expected to be small in the limit. Our next task is to prove that they are both small, and
the proof is built on the great property the matrix A possess: it is usually a lower triangular
matrix, and a strict lower triangular matrix for explicit Runge-Kutta, and thus it is a nilpotent.
Theorem 5.2.2. The method ExpRK-F defined by (5.2.5) is AP for general explicit Runge-
Kutta method with 0 ≤ c1 ≤ c2 ≤ · · · ≤ cν < 1.
Proof. Obviously if R1(λ) = O(ε) and R2(λ) = O(ε) for ε small enough, the theorem holds.
In fact, for explicit Runge-Kutta method, A is a strict lower triangular matrix, and thus a
nilpotent, then one has the following
E−1 (I− CA)−1 E = E−1(I + CA + C2A2 + · · ·+ Cν−1Aν−1
)E (5.2.35a)
= I + B + B2 + · · ·+ Bν−1 (5.2.35b)
where Aν = 0, definition B = CE−1AE and E−1A2E = E−1AEE−1AE are used. According to
the definition of A and E, it can be computed that
Bij = CAijeciλ−cjλ =Cλ
µaij .
Thus I+∑
k Bk is a matrix such that: the element on the kth diagonal is of order O(λk). This
leads to obvious result
R1(λ) = e−λ(
1 +Cλ
µ~b · E−1 (I− A)−1 E · ~e
)= O(e−λλν−1) < O(ε)
Similar analysis can be carried to R2(λ) to show that it vanishes to zero as ε→ 0.
So as ε→ 0, |f l+1− M | → 0. By definition, M is defined by macroscopic quantities computed
directly from the limit Euler equation, thus the numerical scheme is AP, which finishes the
122
proof.
The derivation of the scheme ExpRK-V is essentially the same, and in the end, one still
has, in a condense form
∣∣∣f l+1 −M l+1∣∣∣ ≤ ∣∣∣f l −M l
∣∣∣R1(λ) + ~R2 · ~D (5.2.36)
with R1, ~R2, E, A defined in the same way as in (6.3.7), but Di = |v · ∇xf (i) + ∂tM(i)|.
Following the same computations, one could prove that this method is AP too, but the proof
is omitted for brevity.
Theorem 5.2.3. The method ExpRK-V defined by (5.2.8) is AP for general explicit Runge-
Kutta method.
5.2.3 Numerical example
In this subsection we show three numerical examples. We are especially interested in
checking high order of accuracy and asymptotic preserving.
Convergence rate test
In this example, we use smooth data to check the convergence rate of both methods. The
problem is adopted from [65]: 1 dimensional in x and 2 dimensional in v. Initial distribution
is given by
f(t = 0, x, v) =ρ0(x)
2
(e− |v−u1(x)|2
T0(x) + e|v−u2(x)|2T0(x)
)(5.2.37)
123
with
ρ0(x) =1
2(2 + sin (2πx)) ,
u1(x) = [0.75,−0.75]T , u2(x) = [−0.75, 0.75]T ,
T0(x) =1
20(5 + 2 cos (2πx)) .
Domain is chosen as x ∈ [0, 1] and periodic boundary condition on x is used. Note that the
definition of ρ0, u1/2 and T0 do not represent the real number density, average velocity and
temperature.
As one can see, the initial data is summation of two Gaussian functions centered at u1 and
u2 respectively, and is far away from the Maxwellian. To check the convergence rate, we use
Nx = 128, 256, 512, 1024 grid points on x space, and Nv = 32 points on v space. Time stepping
∆t is chosen to satisfy CFL condition with CFL number being 0.5. We measure the L1 error
of ρ and compute the decay rate through the following formula [199]
error∆x = maxt=tn
‖ρ∆x(t)− ρ2∆x(t)‖1‖ρ2∆x(t)‖1
, (5.2.38)
with ∆x = 1Nx
. The subindex of ρ indicate the discretization on x. Theoretically, a kth order
numerical scheme should give error∆x < C (∆x)k for ∆x small enough.
We compute this problem using spectral method [156] in v, WENO of order 3/5 [190] for x.
For time discretization, we use the second and third order Runge-Kutta from [98], Table 1 page
135. We denote the four schemes under consideration as ExpRK2-F, ExpRK2-V, ExpRK3-F
and ExpRK3-V.
The results for ε = 1, 0.1, 10−3, 10−6 with either Maxwellian initial data or non-Maxwellian
initial data are shown in Figure 5.2.3-5.2.3. We also give the convergence rate Table 5.2.3.
One can see that in kinetic regime, when ε = 1, the two methods are almost the same, but
124
as ε becomes smaller, in the intermediate regime, for example ε = 0.1 for the second order
schemes and ε = 10−3 for second and third order schemes with Maxwellian data, ExpRK-V
performs better then ExpRK-F. In the hydrodynamic regime, however, the two methods give
similar results again shown by the two pictures for ε = 10−6. It is remarkable that the third
order methods achieve almost order 5 (the maximum achievable by the WENO solver) in many
regimes.
−2.8 −2.7 −2.6 −2.5 −2.4 −2.3 −2.2 −2.1 −2−12
−11.5
−11
−10.5
−10
−9.5
−9
−8.5
−8
log10
(dx)
log
10(e
rro
r ρ)
ε=1, Maxwellian Initial
ExpRK2−F
ExpRK2−V
ExpRK3−F
ExpRK3−V
−2.8 −2.7 −2.6 −2.5 −2.4 −2.3 −2.2 −2.1 −2−12
−11.5
−11
−10.5
−10
−9.5
−9
−8.5
−8
log10
(dx)
log
10(e
rro
r ρ)
ε=1, Non−Maxwellian Initial
ExpRK2−F
ExpRK2−V
ExpRK3−F
ExpRK3−V
(a) ε = 1
−2.8 −2.7 −2.6 −2.5 −2.4 −2.3 −2.2 −2.1 −2−12
−11.5
−11
−10.5
−10
−9.5
−9
−8.5
−8
−7.5
−7
log10
(dx)
log
10(e
rro
r ρ)
ε=0.1, Maxwellian Initial
ExpRK2−F
ExpRK2−V
ExpRK3−F
ExpRK3−V
−2.8 −2.7 −2.6 −2.5 −2.4 −2.3 −2.2 −2.1 −2−12
−11.5
−11
−10.5
−10
−9.5
−9
−8.5
−8
−7.5
−7
log10
(dx)
log
10(e
rro
r ρ)
ε=0.1, Non−Maxwellian Initial
ExpRK2−F
ExpRK2−V
ExpRK3−F
ExpRK3−V
(b) ε = 0.1
125
−2.8 −2.7 −2.6 −2.5 −2.4 −2.3 −2.2 −2.1 −2−11
−10
−9
−8
−7
−6
−5
−4
log10
(dx)
log
10(e
rro
r ρ)
ε=10−3
, Maxwellian Initial
ExpRK2−F
ExpRK2−V
ExpRK3−F
ExpRK3−V
−2.8 −2.7 −2.6 −2.5 −2.4 −2.3 −2.2 −2.1 −2−6.5
−6
−5.5
−5
−4.5
−4
−3.5
log10
(dx)
log
10(e
rro
r ρ)
ε=10−3
, Non−Maxwellian Initial
ExpRK2−F
ExpRK2−V
ExpRK3−F
ExpRK3−V
(c) ε = 10−3
−2.8 −2.7 −2.6 −2.5 −2.4 −2.3 −2.2 −2.1 −2−11
−10
−9
−8
−7
−6
−5
−4
log10
(dx)
log
10(e
rro
r ρ)
ε=10−3
, Maxwellian Initial
ExpRK2−F
ExpRK2−V
ExpRK3−F
ExpRK3−V
−2.8 −2.7 −2.6 −2.5 −2.4 −2.3 −2.2 −2.1 −2−6.5
−6
−5.5
−5
−4.5
−4
−3.5
log10
(dx)
log
10(e
rro
r ρ)
ε=10−3
, Non−Maxwellian Initial
ExpRK2−F
ExpRK2−V
ExpRK3−F
ExpRK3−V
(d) ε = 10−6
Figure 5.2.1: ExpRK method. Convergence rate test. In each picture, 4 lines are plotted:the lines with dots, circles, stars and triangles on them are given by results of ExpRK2-F,ExpRK2-V, ExpRK3-F and ExpRK3-V respectively. The left column is for Maxwellian initialdata, and the right column is for initial data away from Maxwellian (5.2.37). Each row, fromthe top to the bottom, shows results of ε = 1, 0.1, 10−3, 10−6 respectively.
126
Initial Distribution Maxwellian Initial Non-Maxwellian InitialNx 128/256/512 256/512/1024 128/256/512 256/512/1024
ε = 1
ExpRK2-F 1.91327 1.99502 1.84968 1.98504ExpRK2-V 2.41608 2.02347 2.67733 2.05436ExpRK3-F 4.99725 4.35014 5.12959 4.76788ExpRK3-V 5.02508 4.40379 5.13515 4.79080
ε = 0.1
ExpRK2-F 1.98218 1.99539 1.97725 1.99454ExpRK2-V 2.41411 2.02293 2.56620 2.05830ExpRK3-F 5.07621 2.94707 5.49587 3.00335ExpRK3-V 5.02220 4.39651 5.13859 4.79264
ε = 10−3
ExpRK2-F 1.23711 1.64976 1.43331 1.73501ExpRK2-V 2.02344 1.85924 1.47466 1.75496ExpRK3-F 2.36140 2.69178 2.55225 2.78275ExpRK3-V 3.86882 3.03223 2.59114 2.80353
ε = 10−6
ExpRK2-F 2.56137 2.04519 2.56137 2.04519ExpRK2-V 2.56137 2.04519 2.56383 2.04859ExpRK3-F 5.08829 4.56695 5.08830 4.56699ExpRK3-V 5.08830 4.56704 4.91909 3.80638
Table 1: Convergence rate for ExpRK methods with different initial data, in different regimes.
A Sod problem
This simple example is adopted from [199] to check accuracy and AP of the numerical
methods. It is a Riemann problem, and the solution to the associated Euler limit is a Sod
problem. (ρ, ux, uy, T ) = (1, 0, 0, 1), if x < 0;
(ρ, ux, uy, T ) = (1/8, 0, 0, 1/4), if x > 0;
As a reference, we compute the problem using forward Euler with dense mesh and small time
step when ε = 0.01: ∆x = 1/500 and h = 0.0001. In Figure 5.2.3 (left), we show that when
ε = 0.01 is comparably big, both methods give the same numerical results with the reference.
In Figure 5.2.3 (right), AP property is shown: it is clear that for ε = 10−6, numerical results
capture the Euler limit – the Euler limit is computed by kinetic scheme [166]. All plots are
given at time t = 0.2.
127
0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 10.1
0.2
0.3
0.4
0.5
0.6
0.7
0.8
0.9
1
x
ρ
ExpRK2−F−Nx100
ExpRK2−V−Nx100
ref: Nx500
0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 10.1
0.2
0.3
0.4
0.5
0.6
0.7
0.8
0.9
1
1.1
x
ρ
ExpRK2−F−Nx100
ExpRK2−V−Nx100
ref: Nx500
(a) ρ
0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1−0.2
0
0.2
0.4
0.6
0.8
1
x
ux
ExpRK2−F−Nx100
ExpRK2−V−Nx100
ref: Nx500
0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1−0.2
0
0.2
0.4
0.6
0.8
1
x
ux
ExpRK2−F−Nx100
ExpRK2−V−Nx100
ref: Nx500
(b) ux
0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1
0.4
0.5
0.6
0.7
0.8
0.9
1
1.1
x
T
ExpRK2−F−Nx100
ExpRK2−V−Nx100
ref: Nx500
0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 10.2
0.3
0.4
0.5
0.6
0.7
0.8
0.9
1
1.1
1.2
x
T
ExpRK2−F−Nx100
ExpRK2−V−Nx100
ref: Nx500
(c) T
Figure 5.2.2: Exp-RK method. The Sod problem. Left column: ε = 0.01. The solid line isgiven by explicit scheme with dense mesh, while dots and circles are given by ExpRK2-F andExpRK2-V respectively, both with Nx = 100. h = ∆x/20 satisfies the CFL condition withCFL number being 0.5. Right column: For ε = 10−6, both methods capture the Euler limit.The solid line is given by the kinetic scheme for the Euler equation, while the dots and circlesare given by ExpRK2-F and ExpRK2-V. They perform well in rarefaction, contact line andshock.
128
−0.5 −0.4 −0.3 −0.2 −0.1 0 0.1 0.2 0.3 0.4 0.50
0.2
0.4
0.6
0.8
1
1.2
x
ε(x)
Kinetic Regime
Euler Regime
Figure 5.2.3: Mixing Regime: ε(x)
Mixing regime
In this example [199], we show numerical results to a problem with mixing regime. This
problem is difficult because ε vary with respect to space. As what we do in the first example,
we take identical data along one space direction, so it is 1D in space but 2D in velocity. An
accurate AP scheme should be able to handle all ε with considerably coarse mesh. Domain is
chosen to be x ∈ [−0.5, 0.5], with ε defined by:
ε =
ε0 + 0.5 (tanh (6− 20x) + tanh (6 + 20x))) x < 0.2;
ε0 x > 0.2
(5.2.39)
where ε0 is 10−3. So ε raise up from 10−3 to O(1), and suddenly drop back to 10−3 as shown
in Figure (5.2.3). Initial data is the give as:
f(t = 0, x, v) =ρ0(x)
4πT0(x)
(e− |v−u0(x)|2
2T0(x) + e− |v+u0(x)|2
2T0(x)
). (5.2.40)
129
with
ρ0(x) = 2+sin (2πx+π)3
u0(x) = 15
cos (2πx+ π)
0
T0(x) = 3+cos (2πx+π)
4
. (5.2.41)
Periodic boundary condition on x is applied. We compute the problem using ExpRK2/3-F/V.
Reference solution is given by Runge-Kutta 2 with fine grids: ∆x = 0.0025.
Results are plotted in Figure 5.2.3. Both methods give excellent results simply taking
a CFL condition of 0.5 whereas explicit methods are forced to operate on a time scale 1000
times smaller. In particular, ExpRK3-V performs well uniformly on ε by giving a more accurate
description of the shock profiles.
130
−0.5 −0.4 −0.3 −0.2 −0.1 0 0.1 0.2 0.3 0.4 0.50.5
0.55
0.6
0.65
0.7
0.75
0.8
0.85
x
ρ
t=0.75
ExpRK2−V−Nx50
ExpRK3−V−Nx50
ref: Nx400
−0.5 −0.4 −0.3 −0.2 −0.1 0 0.1 0.2 0.3 0.4 0.50.5
0.55
0.6
0.65
0.7
0.75
0.8
0.85
x
ρ
t=0.75
ExpRK3−F−Nx50
ExpRK3−V−Nx50
ref: Nx400
(a) ρ
−0.5 −0.4 −0.3 −0.2 −0.1 0 0.1 0.2 0.3 0.4 0.5−0.15
−0.1
−0.05
0
0.05
0.1
x
Ux
t=0.75
ExpRK2−V−Nx50
ExpRK3−V−Nx50
ref: Nx400
−0.5 −0.4 −0.3 −0.2 −0.1 0 0.1 0.2 0.3 0.4 0.5−0.15
−0.1
−0.05
0
0.05
0.1
x
Ux
t=0.75
ExpRK3−F−Nx50
ExpRK3−V−Nx50
ref: Nx400
(b) ux
−0.5 −0.4 −0.3 −0.2 −0.1 0 0.1 0.2 0.3 0.4 0.5
0.65
0.7
0.75
0.8
0.85
0.9
0.95
1
x
T
t=0.75
ExpRK2−V−Nx50
ExpRK3−V−Nx50
ref: Nx400
−0.5 −0.4 −0.3 −0.2 −0.1 0 0.1 0.2 0.3 0.4 0.5
0.65
0.7
0.75
0.8
0.85
0.9
0.95
1
x
T
t=0.75
ExpRK3−F−Nx50
ExpRK3−V−Nx50
ref: Nx400
(c) T
Figure 5.2.4: ExpRK method. The mixing regime problem. The left column shows comparisonof RK2 and RK3 using the ExpRK-V. The solid line is the reference solution with a very finemesh in time and ∆x = 0.0025, the dash line is given by RK3 and the dotted line is givenby RK2, both with Nx = 50 points. The right column compare two methods, both given byRK3, with the reference. The dash line is given by ExpRK-V, and the dotted line is givenby ExpRK-F. Nx = 50 for both. h is chosen to satisfy CFL condition, in our case, the CFLnumber is chosen to be 0.5.
131
Chapter 6
Multi-species systems
In this chapter we tackle the multi-species problem. To analyze the multi-species system
can be very tricky theoretically. The theory carried out for mono-species could not be naıvely
applied over: different particles do not use the same collisional cross-section; they exchange
momentum and energy; particles with different masses converge to the Maxwellian in different
speed, and all these are generating new difficulties on the theoretical level. Despite this, on the
numerical level, we could still explore extensions of the developed computational treatments.
New difficulties are encountered as well, however, most of the currently available methods
mainly only rely on the limits of the systems, which are still known, and thus the difficulties
emerged from collision kernels are all avoided. The arrangement of this chapter is the following:
we review the general properties of the multi-species Boltzmann in section 6.1, and in the
following two sections 6.2 and 6.3, we generalize the two numerical solvers mentioned in the
previous chapter.
6.1 The multi-species system
In this section, we show the multi-species Boltzmann equation and its properties. We also
discuss a simplified BGK model.
132
6.1.1 The multi-species Boltzmann equation
Given a N -species system, we use fi(t, x, v) to represent the distribution function of the
i-th species at time t on the phase space (x, v), and f = (f1, f2, · · · , fN )T is a vector. The
Boltzmann equation for the multi-species system is given by [63]:
∂tfi + v · ∇xfi = Qi(f, f), t ≥ 0, (x, v) ∈ Rd × Rd, (6.1.1)
and the collision operator now is a collection of collisions among all species:
Qi(f, f) =
N∑k=1
Qik(f, f), (6.1.2a)
Qik(f, f)(v) =
∫Sd−1
+
∫Rd
(f ′if′k∗ − fifk∗)Bik(|v − v∗|,Ω)dv∗dΩ. (6.1.2b)
We use the same notation for all the variables, but there are some new changes: the new
collision kernel Bik now depends on both two species i and k. Bik = Bki for symmetry, (thus
Qik = Qki); the computation of the post-collisional velocities v′ and v′∗ also change: the mass
difference should also be taken into account:
v′ = v − 2µikmi
(g · Ω)Ω, v′∗ = v∗ +2µikmk
(g · Ω)Ω, (6.1.3)
with µik = mimkmi+mk
being the reduced mass and mi, mk being the mass for species i and k
respectively. This deduction is based on momentum and energy conservations:
miv +mkv∗ = miv′ +mkv
′∗, mi|v|2 +mk|v∗|2 = mi|v′|2 +mk|v′∗|2.
133
6.1.2 Properties of the multi-species Boltzmann equation
In d-dimensional space, we define the macroscopic quantities for species i: ni is the number
density; ρi is the mass density (note due to mass differences, mass density and number density
have different values); ui is the average velocity; Ei is the total energy; ei is the specific internal
energy; Ti is the temperature; Si is the stress tensor; and qi is the heat flux vector, given by:
ρi = mini = mi
∫fidv, Ei =
1
2ρiu
2i + niei =
1
2mi
∫|v|2fidv,
ρiui = mi
∫vfidv, ei =
d
2Ti =
mi
2ni
∫fi|v − ui|2dv, (6.1.4)
qi =1
2mi
∫(v − ui)|v − ui|2fidv, Si =
∫(v − ui)⊗ (v − ui)fidv,
We also have global quantities for the mixture: the total mass density ρ, the number density
n, the mean velocity u, the total energy E, the internal energy ne and the mean temperature
T = 2ed are defined by:
ρ =∑i
ρi, n =∑i
ni, ρu =∑i
ρiui, (6.1.5a)
E = ne+1
2ρ|u|2 =
d
2nT +
ρ
2|u|2 =
∑i
Ei. (6.1.5b)
Conservations
In the gas mixture system, for each species, the mass is conserved, but the momentum
and energy are not. Species keep exchanging momentum and energy till equilibrium state is
achieved. Mathematically, while taking moments, one only gets zero from the first moment
of the collision term, and the second and the third moments are nontrivial. Except for very
special cases, usually these moments cannot even be explicitly obtained. We list here the
134
explicit expressions of the moments of the Maxwell molecule.
< miQi > =∫miQi(f)dv = 0
< mivQi > =∫mivQi(f)dv =
∑k=1 2µikχiknink[uk − ui]
< 12miv
2Qi > =∫mi2 |v|2Qi(f)dv
=∑
k=1 2miχiknink
[(µikmi
)2 (|uk − ui|2 + 2 ek
mk+ 2 ei
mi
)+ µik
mi
((uk − ui) · ui − 2 ei
mi
)],
(6.1.6)
where χik =∫
(cos θ)2Bik(θ)dθ, with θ = arccos(g·Ω|g|
). One can check [75] for power law
molecules.
Based on these formulas, when taking moments of the Boltzmann equation, one obtains
the corresponding evolution of the macroscopic quantities. Taking the 1D Maxwell molecule
for example:
∂tρi + ∂x(ρiui) =< miQi >= 0, or ∂tni + ∂x(niui) = 0,
∂t(ρiui) + ∂x(Si + ρiu2i ) =
1
ε< mivQi >=
1
ε
∑k
2Bikninkµik[uk − ui], (6.1.7)
∂tEi + ∂x(Eiui + Siui + qi) =1
ε<
1
2mi|v|2Qi >=
1
ε
∑k
2Biknink
(µ2ik
mimk
)(a+ b),
where a = (mkuk +miui)·(uk−ui), b = 2(ek−ei). The macroscopic equations are complicated.
The 1ε terms on the right hand side are also making this set of equation very stiff, not to say
that in most cases this explicit expressions are not even available. Directly designing macro
solver for this set of equation can be very tedious, and is not in our interest. We seek for neat
algorithms that aim at the essential physical meaningful limits: in fact, if we look at the system
as a whole, it will be seen that the total momentum and total energy are still conserved. We
135
sum up the momentum and energy equations for all species and get:
∂t(ρu) + ∂x
(∑i
Si +∑i
ρiu2i
)=
1
ε
∑i
< miviQi >= 0,
∂tE + ∂x
(∑i
Eiui +∑i
Siui +∑i
qi
)=
1
ε
∑i
<1
2miv
2iQi >= 0. (6.1.8)
The conservation hold for all kinds of cross-sections.
The local Maxwellian
As for single species system, we look for the local equilibrium here as well: it is reached
when the gaining part and losing part of all collision terms balance out, namely Qi(f) = 0 for
each i [1]:
fi = M i = ni
( mi
2πT
)d/2e−
mi|v−u|2
2T , (6.1.9)
where T is the mean temperature and u is mean velocity defined in (6.1.5a) and (6.1.5b). We
call this Maxwellian the “unified Maxwellian” because the velocity u and temperature T are
given by those for the entire system instead of those for each single species.
Evidently, we now have requirement on both microscopic level and macroscopic level: the
distribution functions should behave like Maxwellian functions, and the macroscopic physical
observables are required to be the same, namely, average velocity u and temperature T . Also
note that mi appears inside the exponential, indicating that different mass leads to difference
variance of the Gaussian function. In particular, the heavier the particles are, the more likely
they are going to concentrate around u on the velocity space.
136
The Euler limit
The standard Chapman-Enskog expansion is still valid. We expand fi around the unified
Maxwellian (6.1.9), the collision term vanishes, and the system yields its Euler limit [1]:
∂tρi +∇ · (ρiu) = 0,
∂t(ρu) +∇ · (ρu⊗ u+ nT I) = 0, (6.1.10)
∂tE +∇ · ((E + nT )u) = 0.
Here I is the identity matrix. Note that in the equation for ρi, we have u instead of ui as in
(6.1.7). This is because when ε→ 0, ui → u and Ti → T for all i.
6.1.3 A BGK model
In history, many BGK models have been proposed to simulate the multi-species Boltzmann
equation. Unlike the situation in single-species system, here the definition of the BGK operator
is very vague. For a period, people were not sure on whether to design one BGK term for each
single collision operator, or for their summation. However, several criteria in designing BGK
models are clear:
• positivity: the distribution function should be positive;
• indifferentiability: when different species share the same mass, equations of the system
should be consistent with the single species Boltzmann equation.
Some models suffer from the loss of positivity [72], while others fail to satisfy the indifferentia-
bility principle [88]. Here we present one model proposed by Andries, Aoki and Perthame [1]
that meets both standards.
137
The model is:
∂tfi + v · ∇xfi =νiε
(Mi − fi), (6.1.11)
where νi is the collision frequency and Mi is a Maxwellian function:
νi =∑k
nkχik, Mi = ni
(mi
2πTi
)d/2e−mi|v−ui|
2
2Ti . (6.1.12)
The way M is defined is to capture the moments of the collision Q, i.e. νi
(Mi − fi
)shares
the same first d + 2 moments as Qi. As stated above, usually explicit expression is hard to
obtain for the moments of the collision term, and here we only write down that for the Maxwell
molecule (6.1.6):
νiρiui − νiρiui =< mivQi >=n∑k=1
2µikχiknink[uk − ui], (6.1.13a)
νiEi − νiEi
= <1
2miv
2Qi > (6.1.13b)
=n∑k=1
2miχiknink
[(µikmi
)2(|uk − ui|2 + 2
ekmk
+ 2eimi
)+µikmi
((uk − ui) · ui − 2
eimi
)],
and Ti =(
2Ei − ρiu2i
)/(nid). Note that in this case the right hand side of equation (6.1.13a)
is just a linear operator of macroscopic velocities. Also, given known u, the right hand side of
(6.1.13b) is linear on e. To write it in a condense form, we define a matrix L:
(L)ij =
2µijχijninj , i 6= j,
−2∑
k µikχiknink, i = j.
(6.1.14)
138
It is easily seen that under this definition, (6.1.13a) is:
νiρiui − νiρiui = Lu. (6.1.15)
Note that L is a symmetric matrix with each row summing up to 0, and all non-diagonal entries
are positive. Since L is a symmetric weakly diagonally dominant matrix, it is semi-negative
definite, i.e. all its eigenvalues are non-positive.
We also mention another type of Maxwellian, which is defined by macroscopic quantities
ui and Ti for each species. We call it the “species Maxwellian”:
Mi = ni
(mi
2πTi
)d/2e−mi|v−ui|
2
2Ti . (6.1.16)
Remark 6.1.1. Mi− fi can not be used as a BGK operator. In the multi-species system, one
has to introduce some mechanism into the collision term that captures the interactions between
species. Mi − fi gives no communication between the species, so it cannot be used to express
the multi-species collision.
6.2 BGK-penalization method
We generalize the BGK-penalization method [65] to compute the multi-species system.
We show the numerical method in subsection 6.2.1, its properties in 6.2.2, and the numerical
examples can be found in 6.2.4. In subsection 6.2.3, we deal with a special case when the two
particles have disparate masses.
139
6.2.1 Numerical method
Here we use the strategy discovered in [65] and write our scheme as:
f l+1i − f lih
+ v · ∇xf li =Qi(f
l)− Pi(f l)ε
+Pi(f
l+1)
ε. (6.2.1)
The superscript l stands for the time step and h is time discretization. We also use shorthands
Qli , Qi(fl), P li , Pi(f
l) for convenience. Pi is chosen to be the BGK operator:
Pi = β(M i − fi), (6.2.2)
where β is a positive constant carefully chosen to take care of stability and positivity, to be
explained later. A simple algebraic manipulation on gives:
f l+1i =
εf li + h(Qli − βl(Mli − f li ))− εhv · ∇xf li + βl+1∆tM
l+1i
ε+ βl+1h. (6.2.3)
We show the computation of M i, selection of β, computation of the collision term Qi and the
flux term v · ∇xf in order.
The computation of Ml+1
The selection of M i is very crucial. As will be proved later, the distribution function f is
forced to M i algebraically. So the best choice for M i would be simply take the asymptotic
limit of the distribution function f , which is the unified Maxwellian function defined in (6.1.9).
We leave the AP proof to the next subsection. In order to update f , we need to find M i in the
next time step, and thus a simple solver for all its associated macroscopic quantities including
ρi, u and T . To do that, we make use of the fact that these quantities are for the entire system
instead of single species, and are governed by the limiting Euler equation (6.1.10). Instead of
140
directly solving this equation, we use the kinetic scheme by taking the moments of (6.2.1):
nl+1i = nli − h
∫v · ∇xf li dv,
(ρu)l+1 = (ρu)l − h∑i
mi
∫v ⊗ v∇xf li dv,
El+1 = El − h∑i
∫mi
2|v|2v · ∇xf li dv,
and T follows naturally from (6.1.5b):
T l+1 =2El+1 − (ρu2)l+1
dnl+1.
The collision term Q
The computation of collision terms for multi-species systems is much more tricker that in
single-species case. We desire spectral accuracy, and adopt the spectral method introduced
in [165]. To the our best knowledge, this hard problem has never been addressed in any
literature, and the results presented here is only partial results obtained in [112], and though
the idea works for many types of particles, we have access to explicit expression only for the
simplest case: 1D Maxwell molecule. Here we present the computation of its Qik.
Use a ball B(0, S) to represent the domain of the compactly supported distribution f . We
periodize f on v ∈ [−L,L]d with L ≥ (3 +√
2)S. L is chosen much larger than S to avoid
non-physical collision at different periods of the periodized f . Define the Fourier transform as:
f(x; p) =
∫f(x; v)e−ip·v dv, f(x; v) =
1
(2L)d
∑p
f(x; p)eik·v. (6.2.4)
141
Plugging into the collision term (6.1.2b):
Qik =
∫ ∫Bik
[f ′if′k∗ − fifk∗
]dv∗dΩ ≡ Q+
ik − fiQ−ik,
where we define the gaining part and losing part as:
Q+ik =
∫ ∫Bik
(f ′if′k∗)dv∗dΩ, fiQ
−ik =
∫ ∫Bik (fifk∗) dv∗dΩ = fi
∫ ∫Bikfk∗dv∗dΩ.
Using the Fourier transform (6.2.4), one gets:
Q+ik =
∫ ∫Bik
(2L)2d
[∑p
∑q
fi(x; p)eip·v′fk(x; q)eiq·v
′∗
]dv∗dΩ.
For easier computation, one can rewrite equations (6.1.3) to:
v′ = v − µikmi
(g − |g|ω), v′∗ = v∗ +µikmk
(g − |g|ω) = v − g +µikmk
(g − |g|ω).
Note the domain for ω is the entire unit sphere Sd−1 instead of the semi-sphere for Ω. Then:
Q+ik =
1
(2L)2d
∑p,q
fpi fqk
∫ ∫Bike
i(p·v′+q·v′∗)dv∗dω
=1
(2L)2d
∑p,q
fpi fqkei(p+q)·v
∫ ∫Bike
iλ·g+i|g|γ·ωdv∗dω,
where λ = −mkmi+mk
p + −mkmi+mk
q and γ = mkmi+mk
p − mimi+mk
q. Given a specific Bik one can
analytically compute the integration above. The expression, however, can be very tedious,
even in one dimensional space, especially when Bik depends on |g|. Note that the integration
domain for g is not symmetric. For the 1D Maxwell molecule, Bik is a constant, and can be
142
pulled out of the integral, making the computation much easier. In this case,
v′ = v − 2mk
mi +mk(v − v∗) =
mi −mk
mi +mkv +
2mk
mi +mkv∗,
v′∗ = v +mi −mk
mi +mk(v − v∗) =
2mi
mi +mkv − mi −mk
mi +mkv∗.
Plugging in Q+ik, one gets:
Q+ik =
Bik(2L)2
∑p,q
fpi fqkei(mi−mkmi+mk
p+2mi
mi+mkq)v∫ei(
2mkmi+mk
p−mi−mkmi+mk
q)v∗dv∗.
One can also write Q+ik as a summation of its Fourier modes Q+
ik(v) = 1(2L)2
∑l Q
l+ik e
ilv where
Ql+ik =
∫Q+ike−ilvdv
=Bik
(2L)2
∑p,q
fpi fqk
∫ei(mi−mkmi+mk
p+2mi
mi+mkq−l)vdv
∫ei(
2mkmi+mk
p−mi−mkmi+mk
q)v∗dv∗
= Bik∑p,q
fpi fqk sinc(a) sinc(b),
where a = (mi−mkmi+mkp+ 2mi
mi+mkq− l)L, and b = ( 2mk
mi+mkp− mi−mk
mi+mkq)L. The FFT and the inverse
FFT are used to speed up the computation.
The computation for fiQ−ik is much simpler in this special case:
fiQ−ik = fi
∫Bikfkdvk = finkBik.
The choice of the free parameter β
β should be chosen as the maximum value of the Frechet derivative ∇Qi(f) [65]. Numeri-
cally, for positivity, we could find its minimum value. We split the collision Q into the gaining
part and the losing part Qi = Q+i − fiQ−i by Q+
i =∑
kQ+ik and fiQ
−i = fi
∑kQ−ik. Plug back
143
in the scheme (6.2.3), one can rearrange the scheme:
f l+1i =
ε(f li − hv · ∇xf li
)+ hQ+
i (f l) +[(βl −Q−i (f l)
)f li + βl+1M
l+1i − βlM l
i
]h
ε+ βl+1h.
To guarantee positivity, it is sufficient to require the followings for all i [199]:
βl > Q−i (f l), βlMli > βl−1M
l−1i .
The flux term v · ∇xfi
Here we give the numerical flux in 1D. Multi-D could be dealt with in the same fashion.
Use v∂xfi,j to denote the flux term for species i at the grid point xj . A shock-capturing finite
volume method we use is [141]:
v∂xfi,j = ν(fi,j1 − fi,j1−1)− ∆x
2ν(sgn(ν)− ν)(σi,j1 − σi,j1−1), (6.2.5)
where ν = v∆x , ∆x is the mesh size. j1 is chosen to be j for v > 0 and j + 1 for v < 0.
σi,j =fi,j+1−fi,j
∆x φi,j where φi,j is the slope limiter. For the van Leer limiter, it takes value as
φ(θ) = θ+|θ|θ+1 and θi,j =
fi,j−fi,j−1
fi,j+1−fi,j reflects the smoothness around grid point xj .
6.2.2 The AP property of the time discretization
We prove asymptotic preserving in this subsection. This method is not strong-AP, but we
show proof on weak-AP and relaxed-AP. We firstly revisit the time discrete scheme (6.2.1):
f l+1i − f lih
+ v · ∇xf li =Qli − β(M
li − f li )
ε+β(M
l+1i − f l+1
i )
ε. (6.2.6)
144
We will show in order that this method is weakly-AP for the Maxwell molecule, and relaxed-
AP for the BGK model given in subsection 6.1.3. AP is trivial for O(1) ε, and we only prove
it when h ε.
In the proof, we are going to use the following notation: L defined in (6.1.14); λ(M) is used
to describe the spectral radius of matrix M; δuli and δT li are variation of ui and Ti from u and
T respectively; D is a diagonal matrix with its elements being ρis.
λ(M) = supk
(|λk(M)|), δuli = uli − ul, δT li = T li − T l,
D = diagρ1, ρ2, · · · , ρN,
where λk(M) are eigenvalues of M.
Weakly-AP
Lemma 6.2.1. For the Maxwell molecule, if δuli = O(ε) and δT li = O(ε) for ∀i, then δul+1i =
O(ε) and δT l+1i = O(ε).
Proof. Rewrite scheme (6.2.6) as:
f l+1i −M l+1
i =ε(−M l+1
i +Mli)− εhv · ∇xf li
ε+ βh+
hQliε+ βh
−(M
li − f li
). (6.2.7)
Take the first moment on both sides. On the left hand side, one gets (ρiui)l+1− (ρiu)l+1, while
on the right hand side, the first term is O(ε). The second term gives:
h
ε+ βh< mivQ
li > =
h
ε+ βh
∑k
2χikµiknink[ulk − uli] (6.2.8)
=h
ε+ βh
∑k
2χikµiknink(δulk − δuli) = O(ε).
145
The third term gives:
< miv(Mli − f li ) >= ρi
(ul − uli
)= O(ε).
So the entire right hand side is ofO(ε), thus the term on the left hand side, (ρiui)l+1−(ρiu)l+1 =
O(ε), i.e. δul+1i = O(ε). Similar analysis can be carried out for δT .
Remark 6.2.1. In the proof, we used that the collision kernel is for the Maxwell molecule
in (6.2.8). The proof can be extended to other collision kernels as long as one can show the
moments of Q is of O(ε) whenever δu = O(ε), which is usually the case.
Theorem 6.2.1. The method is weakly-AP; namely, if Mli − f li = O(ε), then M
l+1i − f l+1
i =
O(ε).
Proof. Since Mli − f li = O(ε), both Pi(f
l) and Qi(fl) are of O(ε). Plugging back into the
scheme (6.2.7), one gets f l+1i −M l+1
i = O(ε).
Relaxed-AP
Lemma 6.2.2. For the Maxwell molecules, when h 1, in the limit of ε→ 0, there ∃L, such
that ∀l > L, δul = O(ε), given big enough β.
Proof. We prove the result for the 1D case. The proof for higher dimension can be carried out
similarly. The proof follows that of [65]. Take moments of numerical scheme (6.2.6), one gets:
(ρu)l+1 − (ρu)l
h+ ∂x
∫v2mf ldv =
1
ε(Llul + βlDlδul − βl+1Dl+1δul+1).
146
Simple algebra gives:
(ε+ βl+1h
)Dl+1δul+1 =
((ε+ βlh)Dl + hLl
)δul
+ ε(
(ρu)l − (ρu)l+1)− εh∂x
∫v2mf ldv,
⇒ (ε+ βl+1h)(Dl +O(h)
)δul+1 =
[(ε+ βlh)Dl + hLl
]δul +O(ε).
In the derivation, we used Llu = 0 and Dl+1 = Dl + O(h). Invert the coefficients on the left
hand side, one gets:
δul+1 =
[βl
βl+1I +
1
βl+1(Dl)−1Ll +O(h)
]δul +O(ε).
Since the eigenvalues for L are non-positive, if one chooses βl+1 +βl > λ((Dl)−1Ll), given small
enough h, the spectrum of the big bracket on the left hand side is controlled by 1, and thus in
the limit of ε→ 1, δu would decrease to O(ε), and we get our conclusion.
The same analysis can be carried out for T . We call this property proved above “macro-AP”.
It is a property only exists in multi-species problems.
Remark 6.2.2. The proof above is also valid for the BGK model in subsection 6.1.3. It can be
extended to other kinds of collision kernel as well, but the corresponding L may not be linear
on u, and we cannot find explicit expression about the requirement on β. In that case, λ(L)
stands for the spectrum of the nonlinear operator L.
Theorem 6.2.2. If the problem is macro-AP, then, ∃L such that ∀l > L, M li −M
li = O(ε),
and M li −M
li = O(ε).
Proof. It is a straightforward conclusion from the lemma above, and from the definition for M
in (6.1.12).
147
Remark 6.2.3. Up to now, we have shown that Mi approaches to M i for the Boltzmann
collision operator with the Maxwell molecule collision kernel. Rearranging scheme (6.2.6), one
gets:
f l+1 −M l+1=
(ε+ βh)(f l −M l) + hQl
ε+ βh+ε(M
l −M l+1)− εh v · ∇xf l
ε+ βh. (6.2.9)
The second term on the right is of O(ε). So, one can get relaxed-AP only if it can be shown
that Q and f −M have opposite signs. We can prove this for limited form of Q, say the BGK
operator introduced in subsection 6.1.3. However, as will be seen in section 6.2.4, numerically
the scheme is relaxed-AP for the general Boltzmann collision.
Theorem 6.2.3. The scheme is relaxed-AP for the BGK operator Q = ν(M − f) defined in
section 2.3.
Proof. Plug in the definition for Q, (6.2.9) writes:
f l+1 −M l+1=ε+ βh− νhε+ βh
(f l −M l) +
νh(M l −M l)
ε+ βh+O(ε)
=ε+ βh− νhε+ βh
(f l −M l) +O(ε).
The convergence rate to the unified Maxwellian M is apparently:
α =ε+ βh− νhε+ βh
, (6.2.10)
In the limit of ε → 0, if one has β > ν2 , then |α| < 1, thus |f −M | keeps diminishing until
reaching to O(ε), and we get the relaxed-AP.
148
6.2.3 Disparate masses
This section is for the system of gas mixture with disparate masses in the homogeneous case,
i.e. data given is uniform on physical domain. The mathematical problem was first pointed
out by Grad [86], and has attracted great interests since then. The fundamental example is
plasma, for which, the basic derivation can be found in [177, 25]. For these systems, it is the
different time scalings for different species to reach to the equilibria that makes the problem
difficult. Generally speaking, the light species should be able to get to the equilibrium faster,
that is to say there is a time period when the light species is in hydrodynamic regime while the
heavy species is, on the other hand, in kinetic regime. Analyses of the scalings of the collision
operators have been done based on both postulate physical consideration [39, 126] and formal
derivation [48, 49].
Theoretical rescaling analysis
In homogeneous space, the disparate masses system should be written as:
∂tfH = QH = QHH +QHL =
∫BHH(f ′Hf
′H∗ − fHfH∗) dv∗ +
∫BHL(f ′Hf
′L∗ − fHfL∗) dv∗,
∂tfL = QL = QLL +QLH =∫BLL(f ′Lf
′L∗ − fLfL∗) dv∗ +
∫BLH(f ′H∗f
′L − fH∗fL) dv∗.
Now the small parameter that makes the collision term stiff is the ratio of mass ε =√mL/mH
where the sub-indices H and L stand for heavy and light respectively. While assuming that the
two species have densities and temperatures of the same order of magnitude, one could obtain
that fH is much narrower than fL as shown in Figure 6.2.1. To analyze the magnitude of the
collision terms, we define fH(v) = fH(εv) to stretch fH to a function that has comparable
variance as fL. As derived in [48, 49], the scaling ratio of the two collision terms is QH/QL =
O(ε), which means that the collision QL has stronger effect, and that the light species gets to
149
the hydrodynamical regime much faster. For convenience, we write both QH and QL as O(1)
Figure 6.2.1: Distribution function for disparate masses system. The distribution function forheavy species is much narrower than that of the light species.
term, and put ε in front of QH to represent its magnitude. The system turns out to be:
∂tfH = εQH , ∂tfL = QL. (6.2.11)
One can also rescale the time by τ and obtain:
∂tfH =ε
τQH , ∂tfL =
1
τQL. (6.2.12)
When τ = O(ε), the light species is in hydrodynamic regime but the heavy one is still in kinetic
regime; and when τ = O(ε2), both species should be close to the equilibria.
Remark 6.2.4. The inhomogeneous problem gets even harder to analyze, especially when the
different species have different spatial rescaling coefficients. But numerically it makes very
little difference: one simply needs to add the flux term v · ∇xf to the homogeneous scheme.
150
The numerical scheme
The scheme we adopt for (6.2.11) is:
f l+1H − f lH
∆t=ε
τ
(QlH − β
(M lH − f lH
))+βε
τ
(M l+1H − f l+1
H
), (6.2.13a)
f l+1L − f lL
∆t=
1
τ
(QlL − β
(M lL − f lL
))+β
τ
(M l+1L − f l+1
L
), (6.2.13b)
where β = O(1).
Theorem 6.2.4. This scheme yields the following behavior at: O(1ε ) and O( 1
ε2).
• at τ = O(1ε ), the scheme is first order consistent to ∂tfH = QH , and f lL is an O(ε)
approximation of ML;
• at τ = O( 1ε2
), both f lH and f lL are within O(ε) of the unified Maxwellians MH and ML
respectively.
Proof. To show the second statement:
At this time scale, τ = O(ε2), the system turns out to be:
∂tfH =1
εQH , ∂tfL =
1
ε2QL.
By the same arguments as in the previous sections, one gets:
f lH −MlH = O(ε), f lL −M
lL = O(ε2).
for l large enough.
151
To show the first statement:
At this time scale, τ = O(ε), system (6.2.12) can be written as:
∂tfH = QH , ∂tfL =1
εQL.
The scheme still gives f lL →MlL. One just needs to show that the scheme gives a correct
discretization of the equation for fH . Write the scheme as (set τ = ε):
f l+1H − f lH
h= QlH − β(M l
H − f lH) + β(M l+1H − f l+1
H ).
By pulling the fH on the right hand side to the left, one gets:
f l+1H − f lH
h= QlH −
βh
1 + βhQlH +
β
1 + βh
(M l+1H −M l
H
).
The second and the third terms on the right are both O(h), i.e. the scheme gives a first
order discretization to ∂tfH = QH .
6.2.4 Numerical examples
For comparison, the examples chosen are similar to those in [119]. We also perturb the
data on the level of macroscopic quantities. For all the examples below: when ε is not very
small so that solving the Boltzmann equation is still possible by using the basic explicit scheme
with a resolved mesh, we compute the reference solution using forward Euler with very fine
mesh, and compare our results with that; when ε is unbearably small for the forward Euler, we
compare our results to the Euler limit. The Euler equations are computed using CLAWPACK
Euler solver [142].
152
A stationary shock
In this example, we show numerical solution to a Riemann problem of two species. The
analytical solution to the Euler equations is a stationary shock. Here the subscripts 1 and 2
stand for different species.
m1 = 1,m2 = 1.5, n1 = n2 = 1, u1 = 1.8, u2 = 1.3, T1 = 0.3, T2 = 0.35, if x < 0;
m1 = 1,m2 = 1.5, n1 = n2 = 1.401869, u1 = u2 = 1.07, T1 = T2 = 0.8605, if x > 0.
(6.2.14)
The initial distribution for f is given by summation of two Gaussian functions, so it is far away
from the unified Maxwellian M :
fi(t = 0, x, v) =2∑`=1
Ai,` exp(−Bi,`(v − Ci,`)2
), i = 1, 2, (6.2.15)
with
Bi,` =ρi
4Ei − 2ρiu2i (1 + κ2)
, Ai,` =ni2
√Bi,`π, Ci,1 − ui = ui − Ci,2 = κui (6.2.16)
i.e. the two Gaussian functions have the same height and variation, but their centers are 2κu
away from each other. In the numerical experiment, we choose κ = 0.2, ∆x = 10−2 and h is
chosen to satisfy the CFL condition: 10−3 in our simulation. Numerically we check whether
the scheme gives the Euler limit when ε goes to zero; and whether it matches well with the
forward Euler method with relatively fine mesh when ε is big. We first show in Figure 6.2.2
that as ε goes to zero, the numerical solution converges to the Euler limit, the stationary shock
in this case. In Figure 6.2.3, we show that the AP scheme matches very well with the numerical
results given by the forward Euler method for ε = 10−1. Then we show in Figure 6.2.4, that
given an initial data far away from the unified Maxwellian M , f gets close to M quickly with
153
ε = 10−5. This verifies that the scheme is relaxed-AP numerically. Figure 6.2.5 shows that
smaller ε gives faster convergence to the equilibria for macroscopic quantities.
A Sod problem
In this example, we solve a Sod problem. The initial data are given by:
m1 = m2 = 1, n1 = 1, n2 = 1.2, u1 = 0.6, u2 = −0.5, T1 = T2 = 0.709, if x < 0;
m1 = m2 = 1, n1 = 0.125, n2 = 0.2, u1 = −0.2, u2 = 0.125, T1 = T2 = 0.075, if x > 0.
(6.2.17)
The initial distribution is chosen far away from the Maxwellian as defined in (6.2.15) and
(6.2.16) with κ again chosen as 0.2. For all ε, we choose ∆x = 10−2 and h = 10−3. In this
problem, m1 = m2, so we first show the numerical indifferentiability in Figure 6.2.6, that
is: computing the problem as a multispecies system gives the same result as computing the
monospecies Boltzmann equation. In Figure 6.2.7, we show that as ε goes to zero, the numerical
solution converges to the Euler limit. For ε as big as 10−1 and 10−2, we compare the results
with those of the forward Euler with a fine mesh. They match well as shown in Figure 6.2.8.
In Figure 6.2.9, we show that for ε = 10−4, although the initial f is far away from the unified
Maxwellian M , as time evolves, it converges to M . This numerically verifies the relaxed-AP
property. In Figure 6.2.10, we show the evolution of u with different ε. Apparently different
species gradually share the same velocity, and the smaller ε is, the faster the convergence is.
A disparate masses problem
In this example, we solve the disparate masses system. Define ε =√
mLmH
, we want to verify
that the light species gets close to the unified Maxwellian ML faster than the heavy one. We
154
solve an inhomogeneous problem with the following initial data:
mH = 8, mL = 0.08, uH = 0, uL = 0.5, TH = TL = 2.5, nH = 1, nL = 1.2.
The initial distribution functions are still given by the summation of two Gaussian functions
as in (6.2.15), with parameters A1 = A2 and B1 = B2 defined in (6.2.16), and C1 and C2
defined by: C1 − u = u− C2 = κ. We choose κ = 0.5 for the heavy species and κ = 4 for the
light one. In Figure 6.2.11, on the left we show the initial distribution functions for the two
species, both of which are given by summation of two Gaussian functions and are far away
from the Maxwellian. On the right we show several snapshots of the distribution functions as
they evolve. In Figure 6.2.12, we show that as time evolves, the velocities converge toward
each other. Note that the heavy species weighted more when computing for the mean velocity
u as in (6.1.5a), thus its average velocity does not change much.
6.3 ExpRK method for the multi-species Boltzmann equation
In this section we explore how to extend the Exponential Runge-Kutta method to multi-
species system [144]. Parts of the computation is the same as in the previous section and will be
omitted. We present the method in subsection 6.3.1 and prove its properties in subsection 6.3.2.
Numerical examples will be give in subsection 6.3.3.
6.3.1 Exponential Runge-Kutta method
It is a simple extension of the original Exponential Runge-Kutta method [52, 143]. One
encounters the same difficulty as in the BGK-penalization case. The algorithm consists two
parts in each time step:
155
Figure 6.2.2: Multi-species Boltzmann equation. BKG-penalization method. The stationaryshock problem. As ε→ 0, solution of the Boltzmann equation goes to the Euler limit. t = 0.1.
156
Figure 6.2.3: Multi-species Boltzmann equation. BKG-penalization method. The stationaryshock problem. ε = 0.1, t = 0.1. The dashed line is given by the AP scheme, and the solidline is given by the forward Euler with a fine mesh: ∆x = 0.01 and h = 0.0005.
Figure 6.2.4: The stationary shock problem. ε = 10−5, δf = f1 − M1 diminishes as timeevolves.
157
Figure 6.2.5: Multi-species Boltzmann equation. BKG-penalization method. The stationaryshock problem. u1 and u2 at x = −0.5 on the left and T1, T2 on the right, as functions of time,for ε = 10−2 and 10−5 respectively. Note the different time scales for the two figures.
t=0.1
−0.25 −0.2 −0.15 −0.1 −0.05 0 0.05 0.1 0.15 0.2 0.250
1
2
3
x
ρ
as 2 species
as 1 species
−0.25 −0.2 −0.15 −0.1 −0.05 0 0.05 0.1 0.15 0.2 0.25−0.5
0
0.5
1
1.5
x
U
as 2 species
as 1 species
−0.25 −0.2 −0.15 −0.1 −0.05 0 0.05 0.1 0.15 0.2 0.250
1
2
x
T
as 2 species
as 1 species
Figure 6.2.6: Multi-species Boltzmann equation. BKG-penalization method. The Sod problem.Indifferentiability. t = 0.1. ρ, u and T are computed using two species model (“o”) and onespecies model (“.”).
158
Figure 6.2.7: Multi-species Boltzmann equation. BKG-penalization method. The Sod problem.t = 0.1. As ε→ 0, the numerical results go to the Euler limit. Solid lines give the Euler limitcomputed by using [142].
159
Figure 6.2.8: Multi-species Boltzmann equation. BKG-penalization method. The Sod problem.t = 0.1. For ε = 0.1, 0.01, we compare the results of the AP scheme, given by the circled lines,and the results of the forward Euler with a fine mesh, given by the solid lines.
Figure 6.2.9: Multi-species Boltzmann equation. BKG-penalization method. The Sod problem.ε = 10−4. δf = f1 −M1.
160
Figure 6.2.10: Multi-species Boltzmann equation. BKG-penalization method. The Sod prob-lem. The three figures show velocities u1 and u2 at x = −0.3 as functions of time forε = 0.1, 0.01, and 10−5 respectively. Note different time scales for three figures.
Figure 6.2.11: Multi-species Boltzmann equation. BKG-penalization method. The disparatemasses problem. The left figure shows the initial distributions and the right figure show thetime evolutions of fH and fL. fH is put at the top and fL is at the bottom. At t = 0.007, fLis close to ML while fH is still far away from the equilibrium. Note the different scales for v.
161
Figure 6.2.12: Multi-species Boltzmann equation. BKG-penalization method. The disparatemasses problem. The velocities for the two species converge to each other.
Step 1. Solve the convection part:
∂tfi + v · ∇xfi = 0. (6.3.1)
Step 2. Solve the collision part:
∂tfi =Qiε. (6.3.2)
This is based on the Strang splitting algorithm, and thus the accuracy is second order in
time. The convection part is easy. We use the same algorithm presented in section 6.2.1 to deal
with the transport term. The collision term, however, is the stiff part: if explicit methods are
used, the time step is limited by ε, and if implicit methods are used, one meets the difficulties
in numerically inverting Q. Following [52, 143], we reformulate (6.3.2) into an exponential
form, and solves the new equation using explicit Runge-Kutta scheme.
Reformulation
Define
µ = supx,i|Q−i |, Pi = Qi + µfi, (6.3.3)
162
and rewrite (6.3.2) as:
∂t((fi −M i)e
µt/ε)
= ∂tfieµt/ε + (fi −M i)
µ
εeµt/ε =
1
ε(Qi + µfi − µM i)e
µt/ε
=1
ε(Pi − µM i)e
µt/ε, (6.3.4)
in which we have used the fact that ρi, u and T do not change in the collision step, and thus
∂tM i = 0.
Remark 6.3.1.
1. The equation (6.3.4) describes the evolution of the distance between the distribution func-
tion fi and its equilibrium M i multiplied by an integration factor, and thus removes the
stiffness in (6.3.2) ([52]).
2. Equation (6.3.4) holds for arbitrary constant µ. The way we choose µ is to guarantee the
positivity of P ; we refer to [143] for more details.
Runge-Kutta method
Applying the K-stage Runge-Kutta method to the equation (6.3.4), one gets
Stage α: (f l,(α) −M)e
µεcαh = (f l+
12 −M) +
h
ε
α−1∑β=1
aαβ(P l,(β) − µM)eµεcβh,
Final stage: (f l+1 −M)eµεh = (f l+
12 −M) +
h
ε
K∑α=1
bα(P l,(α) − µM)eµεcαh.
(6.3.5)
where∑α−1
β=1 aαβ = cα,∑
α bα = 1, and superscripts l, (β) stand for the estimate of the quan-
tities at t = tl + cβh. µ is given in (6.3.3) and f l+12 is obtained after the transport step. The
last equation implies:
f l+1 =
(1− eλ −
∑α
bαλeλ(−1+cα)
)M + e−λf l+
12 +∑α
bαλeλ(cα−1)P
l,(α)
µ, λ =
µh
ε, (6.3.6)
163
In this formulation, one needs to compute M and P . The computation is the same as that in
section 6.2.1, and will be omitted from here.
6.3.2 Positivity and asymptotic preserving properties
In this section, we discuss the positivity and asymptotic preserving properties of the Exp-
RK method introduced above.
Theorem 6.3.1 (Positivity). The Exp-RK method described by (6.3.1) and (6.3.5) preserves
the positivity property of fi, i.e. there exist h∗ > 0 and µ∗ > 0 such that f l+1 > 0 provided
f l > 0, if 0 < ∆x < h∗ and µ > µ∗.
The proof of Theorem 6.3.1 is essentially the same as in the section for single-species system
5.2.2, and we omit the details here.
In order to prove the AP property, we need the following assumption:
Assumption 6.3.1. The operator P satisfies
‖P (f, f)− P (g, g)‖ . ‖f − g‖.
Remark that this assumption is true for Maxwell molecule in the d2 norm defined in Ps(Rd)
space ([183]). This can be easily seen by a similar argument carried in section 5.2.2. For
completeness we provide it in appendix B.3.
Assume f l and gl are the initial conditions to (6.3.5). Define
d0 = ‖f l − gl‖, dα = ‖f l,(α) − gl,(α)‖, α = 1, · · · ,K,
~d = (d1, d2, · · · , dK)T , ~b = (b1, b2, · · · , bK)T , ~e = (1, 1, · · · , 1)T ,
164
and A is a K ×K strictly lower triangular matrix and E is a diagonal matrix given by
Aαβ =λ
µaαβe
(cβ−cα)λ, β < α, and E = diage−c1λ, e−c2λ, · · · , e−cKλ, (6.3.7)
where aαβ, bα and cα are the coefficients of the Runge-Kutta method in (6.3.5), λ is given in
(6.3.6) and ~e is a K-dimensional vector. Then we have the following contraction lemma.
Lemma 6.3.1. After one time step iteration in (6.3.5), the scheme satisfies:
‖f l+1 − gl+1‖ ≤ R(λ)‖f l − gl‖ with R(λ) = e−λ(
1 +Cλ
µ~bTE−1(I− CA)−1E~e
). (6.3.8)
where I is the identity matrix, and C > 0 is a constant.
Proof. The equation (6.3.5) implies, for α = 1, · · · ,K,
(f l,(α) − gl,(α))ecαλ = (f l − gl) +α−1∑β=1
aαβλ
µecβλ(P
l,(β)f − P l,(β)
g ). (6.3.9)
Using the triangle inequality and Assumption 6.3.1 produces
‖f l,(α) − gl,(α)‖ ≤ ‖f l − gl‖e−cαλ +α−1∑β=1
aαβλ
µe(cβ−cα)λ
(C‖f l,(β) − gl,(β)‖
), (6.3.10)
which implies
~d ≤ d0E~e+ CA~d and ~d ≤ d0 (I− CA)−1 E~e, (6.3.11)
where we have used the fact that A is a strictly lower triangular matrix.
The final step in (6.3.5) yields
(f l+1 − gl+1
)=(f l − gl
)e−λ +
K∑α=1
hbαε
(Pl,(α)f − P l,(α)
g
)e(cα−1)λ, (6.3.12)
165
Then (6.3.11) and Assumption 6.3.1 imply
‖f l+1 − gl+1‖ ≤ d0e−λ +
Cλ
µe−λ~bTE−1~d ≤ d0e
−λ(
1 +Cλ
µ~bTE−1(I− CA)−1E~e
), (6.3.13)
which completes the proof.
Lemma 6.3.2. Under Assumption 6.3.1, ‖f l −M l‖ = O (R(λ)) for each l.
Proof. Take g = M in the previous lemma, one gets
‖f l+1 −M l+1‖ ≤ ‖f l+ 12 −M l+ 1
2 ‖R(λ).
The convection step yields
‖f l+ 12 −M l+ 1
2 ‖ < ‖f l −M l‖+ ∆t‖v · ∇x(f l −M l
)‖+O(∆t) ≤ ‖f l −M l‖+O(∆t).
Combine these two inequalities, we conclude with the assertion of this lemma.
Theorem 6.3.2. The Exp-RK method defined in (6.3.1) and (6.3.5) is strong AP.
Proof. By Lemma 6.3.2, it suffices to prove that R(λ) = O(ε) for small ε. Considering that A
is a strictly lower triangular matrix and is thus a nilpotent, one has
E−1 (I− A)−1 E = E−1(I + A + A2 + · · ·+ AK−1
)E = I + B + B2 + · · ·+ BK−1, (6.3.14)
where B = E−1AE, and we have used AK = 0. Further by (6.3.7), Bαβ = Aαβecαλ−cβλ = λµaαβ.
Thus I +∑
α Bα is a matrix with the elements of at most O(λK). Therefore, when ε 1,
R(λ) = e−λ(
1 +Cλ
µ~b · E−1 (I− A)−1 E · ~e
)= O(e−λλK) < O(ε).
166
This completes the proof.
Remark 6.3.2. Note that in the derivation for (6.3.4), the only requirement on M is that it
does not change with respect to time in the collision part. Analytically, one can replace this M
by arbitrary function that does not has the time variable. However, the associated numerics
simply can not preserve the right asymptotic limit, as indicated in Lemma 6.3.2.
6.3.3 Numerical example
The examples chosen here is the same as in the previous section 6.2.4 for the convenience
of comparisons. We compute the reference solution by the forward Euler method using very
small time step and mesh size.
A stationary shock
We compute the same stationary shock problem as shown in 6.2.4. N = 2 in (4.1.1), and
the initial macroscopic quantities form a shock with zero speed, see (6.2.14). In phase space,
the distribution function is given far away from the Maxwellian as in (6.2.15) and (6.2.16) with
κ = 0.2. We choose h = 0.0005 and ∆x = 0.01 to compute the equation using forward Euler
scheme as a reference result. In Figure 6.3.1, we compare the numerical results given by the
Exp-RK method, the BGK penalization method, and the reference solution for ε = 1 and 0.1.
In Figure 6.3.2, we verified the AP property of the Exp-RK method, i.e. as ε goes to zero, the
numerical results capture the stationary shock. We compare the convergence of the velocities
for two species in Figure 6.3.3, where they gradually converge to the mean velocity u, and the
smaller ε gives the faster convergence rate.
167
−0.5 0 0.50.8
11.21.41.61.8
t=0.1, ε=1
ρ
−0.5 0 0.5
1
1.5
u1
−0.5 0 0.5
0.5
1
T1
x
−0.5 0 0.50.8
11.21.41.61.8
t=0.1, ε=0.1
ρ
−0.5 0 0.5
1
1.5
u1
−0.5 0 0.5
0.5
1
T1
x
Figure 6.3.1: Multi-species Boltzmann equation. ExpRK method. The stationary shock prob-lem. We compare the numerical results at t = 0.1 by the Exp-RK method (the dotted line),the BGK penalization method (the circled line), and the reference solution (the solid line).The left figures are for ε = 1, and the right ones are for ε = 0.1.
A Sod problem
This Sod problem is the same one computed in section 6.2.4. Macroscopic quantities are
chosen as in 6.2.17, and the initial distribution function is taken far away from the Maxwellian.
We check indifferentialbility in Figure 6.3.4, and compare the numerical results given by this
Exponential Runge-Kutta method and the BGK-penalization method, together with reference
solution for ε = 1 and 0.1. AP property is verified in Figure 6.3.6 as we take ε→ 0. Figure 6.3.7
shows that in the long time limit, the two species have the same velocities, indicating the Euler
limit is achieved. Smaller ε gives faster convergence rate. f −M at t = 0.1 for kinetic regime
ε = 0.1 and hydrodynamic regime ε = 10−6 are plotted in Figure 6.3.8 reflecting AP on the
microscopic level.
168
−0.3 −0.2 −0.1 0 0.1 0.2 0.30.9
1
1.1
1.2
1.3
1.4
1.5
1.6
t=0.1, ExpRK−1
ρ1
x
ε=1
ε=0.1
ε=10−2
ε=10−4
ε=10−5
−0.3 −0.2 −0.1 0 0.1 0.2 0.31
1.1
1.2
1.3
1.4
1.5
1.6
1.7
1.8t=0.1, ExpRK−1
u1
x−0.3 −0.2 −0.1 0 0.1 0.2 0.3
0.3
0.4
0.5
0.6
0.7
0.8
0.9
1t=0.1, ExpRK−1
T1
x
Figure 6.3.2: Multi-species Boltzmann equation. ExpRK method. The stationary shock prob-lem. As ε goes to zero, the results get to the Euler limit, i.e. a stationary shock.
0 0.05 0.1
1.3
1.4
1.5
1.6
1.7
1.8
1.9
t
ε=1
u
0 0.05 0.1
1.3
1.4
1.5
1.6
1.7
1.8
1.9
t
ε=0.01
0 0.05 0.1
1.3
1.4
1.5
1.6
1.7
1.8
1.9
t
ε=10−5
Figure 6.3.3: Multi-species Boltzmann equation. ExpRK method. The stationary shock prob-lem. The velocities converge to the mean velocity at different rates for different ε’s. The circlesand dots stand for the velocities of the first and second species respectively. The smaller εgives the faster convergence rate.
169
−0.5 0 0.50
0.5
1
1.5
2
2.5
ρ
x
−0.5 0 0.5−0.5
0
0.5
1
ux
−0.5 0 0.50
0.5
1
1.5
T
x
Figure 6.3.4: Multi-species Boltzmann equation. ExpRK method. The Sod problem. Indistin-guishability. The solid lines are the results of treating the system as single species, and theyagree with the results given by the dashed lines, which are the results of a system that has twospecies with the same mass.
−0.2 −0.1 0 0.1 0.20
0.5
1
1.5
t=0.1, ε=1
ρ1
−0.2 −0.1 0 0.1 0.2−0.5
00.5
11.5
u1
−0.2 −0.1 0 0.1 0.2
0
1
2
E
x
−0.2 −0.1 0 0.1 0.20
0.5
1
1.5
t=0.1, ε=0.1
ρ1
−0.2 −0.1 0 0.1 0.2−0.5
00.5
11.5
u1
−0.2 −0.1 0 0.1 0.2
0
1
2
E
x
Figure 6.3.5: Multi-species Boltzmann equation. ExpRK method. The Sod problem. Att = 0.1, we compare the results give by the Exp-RK method (the dotted line), the BGKpenalization method (the circled line) and the reference solution (the solid line). The leftfigures are for ε = 1, and the right ones are for ε = 0.1.
−0.5 0 0.50
0.5
1
1.5
2
2.5t=0.1, ExpRK−1
ρ
x
ε=1
ε=0.1
ε=10−2
ε=10−4
ε=10−5
Euler limit
−0.5 0 0.5
0
0.5
1
t=0.1, ExpRK−1
u1
x−0.5 0 0.5
0
0.2
0.4
0.6
0.8
1
1.2
t=0.1, ExpRK−1
E
x
Figure 6.3.6: Multi-species Boltzmann equation. ExpRK method. The Sod problem. As εgoes to zero, the numerical results by the Exp-RK method converge to the Euler limit.
170
0 0.05 0.1−0.5
0
0.5
1
t
ε=0.1
u
0 0.05 0.1−0.5
0
0.5
1
t
ε=0.01
0 0.05 0.1−0.5
0
0.5
1
t
ε=10−4
Figure 6.3.7: Multi-species Boltzmann equation. ExpRK method. The Sod problem. Thevelocities converge to the mean velocity at different rates for different ε’s. The circled linestands for species one, and the dotted line stands for species two. The smaller ε gives thefaster convergence rate.
−10
0
10
−0.5
0
0.5−0.2
−0.1
0
0.1
0.2
v
ε=0.1, t=0.1
x −10
0
10
−0.5
0
0.5−4
−2
0
2
x 10−9
v
ε=10−6
, t=0.1
x
Figure 6.3.8: Multi-species Boltzmann equation. ExpRK method. The Sod problem. f −Mfor ε = 1 and ε = 10−6 at t = 0.1.
Appendices
171
172
Appendix A
Quantum system
A.1 Some basic analysis of the semi-classical Liouville systems
To understand the asymptotic behavior of the solution to the Liouville-A system (2.2.19),
as mentioned in section 2.2.3, we look at a simpler model system:
∂tg + ∂xf + b(x)∂pg = 0 ,
∂tf + a(p)∂xf + ∂xg + b(x)∂pf = iεc(p)f ,
g(0, x, p) = gI(x, p), f(0, x, p) = fI(x, p).
(A.1.1)
The initial conditions gI and fI are bounded smooth functions independent on ε, b > 0 and
the set of zeros for c(p): Sc = p : c(p) = 0 is measured zero. It is easy to check that (A.1.1)
is a linear hyperbolic system, and the solutions g and f are bounded uniformly in ε [43].
A.1.1 Weak convergence
We consider the weak limit of the solution of (A.1.1) in this subsection. To do this, we
introduce the inner product 〈 ·, · 〉 as
〈u, v 〉 =
∫ ∞0
∫R2
u(t, x, p)v(t, x, p) dx dp dt.
173
Choose an arbitrary test function h ∈ C∞0 (R+ × R2), take the inner product on both side of
(A.1.1) w.r.t h, one gets
〈 ∂tg, h 〉 − 〈 f, ∂xh 〉+ 〈 b∂pg, h 〉 = 0 ,
〈 f, ∂th 〉+ 〈 af, ∂xh 〉+ 〈 g, ∂xh 〉+ 〈 bf, ∂ph 〉 = − iε〈 cf, h 〉 .
(A.1.2)
The derivatives in the equation of (A.1.2) are acted on the smooth function h, and the left
side is bounded. One gets
〈 cf, h 〉 → 0 as ε→ 0 for all h ∈ C∞0 (R+ × R2).
Given that c is almost surely nonzero, and f is bounded, one gets
f 0 weakly.
Combined with the first equation in (A.1.2), one gets
∂tg + b(x)∂pg 0 weakly.
A.1.2 Strong convergence: for constant b
In these two subsections, we formally prove that before getting close to the crossing region,
c(p) is assumed to be bigger than a constant c0 that is unrelated to ε. In this region, f is
constantly small and controlled by O(ε). This subsection is for the case when the speed on p
direction is a constant: b(x) = β. Along the p-characteristic line p(t) = p0 + βt, one applies
the Fourier transform to the x-variable, and gets:
d
dtf = iR(t)f , (A.1.3)
174
where f(t, η) = ( g(t, η, p(t)), f(t, η, p(t)) )T and
R(t) =
0 −η
−η c(p(t))/ε− ηa(p(t))
.
The two eigenvalues of R(t) are both real, and thus the system above has a bounded solution
satisfying
|g(t, η)|2 + |f(t, η)|2 = |gI(η)|2 + |fI(η)|2.
The equivalence between norms gives:
|g(t, η)|+ |f(t, η)| < C(|gI(η)|+ |fI(η)|).
Adopt it into the solution to (A.1.3), one gets
|∂tg(t, x)| = 1
2π
∣∣∣∣∫Rηf(t, η)eiηxdη
∣∣∣∣ ≤ 1
2π
∫R|η| |f(t, η)|dη
≤ C∫R|η|(|gI(η)|+ |fI(η)|
)dη
∣∣∂2txg(t, x)
∣∣ =1
2π
∣∣∣∣∫Rη2f(t, η)eiηxdη
∣∣∣∣ ≤ 1
2π
∫R|η|2 |f(t, η)|dη
≤ C∫R|η|2
(|gI(η)|+ |fI(η)|
)dη
If the initial conditions gI and fI are smooth enough, and decay fast as η → ∞, one could
easily get that ∂tg(t, x) and ∂2txg(t, x) are both bounded in time independent of ε, and thus g
and ∂xg are slowly varying in time.
Remark A.1.1. In the derivation above, we dropped the p(t)-dependence in the functions for
simplicity. The partial derivative in the t-variable ∂t should be understood as taking along the
p-characteristic line p(t), i.e. ∂tg(t, x) = ∂tg(t, x, p) + β∂pg(t, x, p).
175
Assume that fI(x) ≡ 0, one follows the characteristics of x(t) by solving x(t) = a(t, x) and
gets:
f(t, x(t), p(t)
)= −
∫ t
0exp
(i
ε
∫ t
t−sc (p(τ)) dτ
)∂g
∂x
(t− s, x(t− s), p(t− s)
)ds .
By the assumption, before hitting the crossing region, c(p(τ)) > c0 > 0, then the stationary
phase argument suggests that, given slowly varying ∂xg(t, x, p(t)), f = O(ε).
The observations from the above two subsections suggest that f 0, and before getting
close to the crossing region, f is as small as of O(ε). Based on these arguments, for the
Liouville-A system (2.2.19), and propose the following conjecture: if σ12 and σ21 are initially
zero, then:
Case 1. If p −√ε, then σ12 and σ21 are of o(√ε);
Case 2. If p ∈ [−√ε,√ε], then σ12 and σ21 are of O(√ε), and slowly varying;
Case 3. If p √ε, σ12 and σ21 are highly oscillatory, and converge to 0 weakly.
A.2 The integration method of a simple model system
In this section, we apply the method in (2.3.6) onto a simple model to show stability.
d
dtf(t, p0 + t) = R(p0 + t) f(t, p0 + t),
where f(t, p) = (g(t, p), f(t, p))T , p = p0 + t, and
R(p) =
r11(p) r12(p)
−r12(p) rε22(p)
.
176
with rε22(p) = r22(p)+ iεc(p), while r11 and r22 are purely imaginary, and c(p) real and positive.
r11, r22 and r12 are independent on ε.
Set up mesh as tj = j∆t, and pi = −12 + (i− 1)∆p, with ∆p and ∆t being the mesh size.
Denote gji and f ji as the numerical result at (tj , pi), then (2.3.6) gives
f(t, p(t)) = f ji e∫ t0 r
ε22(pi+τ)dτ − gji r12(pi)
∫ t
0e∫ ts r
ε22(pi+τ)dτds, (A.2.1a)
g(t, p(t)) = gji + g(t, p(t)) r11(p(t)) + r12(pi)
∫ t
0f(tj + s, pi + s) ds (A.2.1b)
Plug (A.2.1a) into (A.2.1b), and evaluate them at (tj+1, pi+1), one obtains:
(1− r11(pi+1)∆t
)gj+1i+1 = gji
(1− |r12(pi)|2
∫ ∆t
0
∫ t
0e∫ ts r
ε22(pi+τ)dτdsdt
)+f ji r12(pi)
∫ ∆t
0e∫ t0 r
ε22(pi+τ)dτdt . (A.2.2)
Written in vector form gives
f j+1i+1 = Mi f
ji , (A.2.3)
with
Mi =
1−|r12(pi)|2∫∆t0
∫ t0 e
∫ ts rε22(pi+τ)dτdsdt
1−r11(pi+1)∆t
r12(pi)∫∆t0 e
∫ t0 rε22(pi+τ)dτdt
1−r11(pi+1)∆t
−r12(pi)∫ ∆t
0 e∫∆ts rε22(pi+τ)dτds e
∫∆t0 rε22(pi+τ)dτ
.
The following quantities in the matrix Mi should be evaluated very accurately:
F0 = e∫∆t0 rε22(pi+τ)dτ ,
F1 =
∫ ∆t
0e∫ t0 r
ε22(pi+τ)dτdt,
F2 =
∫ ∆t
0
∫ t
0e∫ ts r
ε22(pi+τ)dτds dt.
Remark A.2.1. These three quantities only depend on the mesh grid point index i but not the
177
time steps index j, thus they only need to be computed once at the beginning of the computation.
Given ε ∆t, the integrands of F1 and F2 are highly oscillatory, and one can see that
|F1| ∼ O(ε) and |F2| ∼ O(ε2). Simple calculation shows that Mi can be written as
Mi = Ω Mi,
with
Ω = diag
(1
1− r11(pj+1)∆t, 1
), Mi =
1− |r12(pi)|2 F2 r12 F1
−r12 F1 F0
.
With purely imaginary r11 and rε22, it is easy to prove that ‖Ω‖∞ ≤ 1, and ‖Mi‖∞ ≤ (1 +
O(ε∆t)), and thus ‖Mi‖∞ ≤ (1 + O(ε∆t)). This implies asymptotic stability of the scheme
(A.2.3) independent of ε→ 0.
A.3 The derivations of (3.2.18)
In this section, we give the derivations of the semi-classical Liouville system (3.2.18). We
firstly study the Moyal product. From its definition in (3.2.8), one gets:
A#B = Aeiε2
(←−∇x·−→∇p−
←−∇p·−→∇x
)B
=∑n
1
n!
(iε
2
)nA(←−∇x ·
−→∇p −←−∇p ·
−→∇x
)nB
=∑n
n∑k=0
(−1)k
n!
(n
k
)(iε
2
)nA(←−∇x ·
−→∇p
)n−k (←−∇p ·−→∇x
)kB.
178
In particular, if both symbols depend only on one variable (spatial x or phase p), the Moyal
product becomes the ordinary product:
A(x)#B(x) = A(x)B(x), and A(p)#B(p) = A(p)B(p).
With this, one can evaluate the symbol H ′.
H ′ = Θ(x)#H(x,p)#Θ†(x)
= Θ(x)#(U(x)I + V (x)
)#Θ†(x) + Θ(x)#
(p2
2I)
#Θ†(x)
= diagU +
√u2 + v2, U −
√u2 + v2
+ Θ(x)#
(p2
2I)
#Θ†(x)
= Λ(x,p) + iεp · ∇xΘ(x)Θ†(x) +ε2
2∇xΘ(x) · ∇xΘ†(x), (A.3.1)
where Λ = U(x) + ΛV with ΛV defined in (3.1.21). Hence we obtain the equation (3.2.17) in
Section 3.2.2. Here the derivation is based on (3.2.15), (3.2.8), and the following identities:
Θ(x)#
(p2
2I)
=∑n
1
n!
(iε
2
)nΘ(x)
(←−∇x ·−→∇p
)n(p2
2I)
=p2
2Θ +
iε
2p · ∇xΘ +
1
2
(iε
2
)2
Θ(←−∇x ·
−→∇p
)2(
p2
2I)
=p2
2Θ +
iε
2p · ∇xΘ +
1
2
(iε
2
)2
∆xΘ,(p2
2Θ(x)
)#Θ†(x) =
p2
2− iε
2Θ (p · ∇xΘ†) +
1
2
(iε
2
)2
Θ ∆xΘ†,(p · ∇xΘ(x)
)#Θ†(x) = p · ∇xΘ Θ† − iε
2∇xΘ · ∇xΘ†,
∆xΘ(x)#Θ†(x) = ∆xΘ Θ†,
Θ(x) (p · ∇xΘ†(x)) = −(p · ∇xΘ) Θ†,
−2∇xΘ(x) · ∇xΘ†(x) = Θ∆xΘ† + ∆xΘ Θ†.
179
To finally derive (3.2.18), one needs to go back to the evolution of F ′ given in (3.2.14):
iε∂F ′
∂t= [Λ, F ′]− iε
2Λ, F+
iε
2F,Λ+ iε[p · ∇xΘ Θ†, F ′] + O(ε2)
= [Λ, F ′]− iε
2[∇pΛ,∇xF ]+ +
iε
2[∇xΛ,∇pF ]+
+ iε [p · ∇xΘ Θ†, F ′] + O(ε2), (A.3.2)
with [A,B]+ = AB +BA. By definition (3.2.20), it is easy to obtain:
[Λ, F ′] =
0 λ+ − λ−
λ− − λ+ 0
, [∇pΛ,∇xF ]+ = 2
p · ∇xf+ p · ∇xf
i
p · ∇xf i p · ∇xf−
,
[∇xΛ,∇pF ]+ = 2
∇x(U + E) · ∇pf+ ∇xU · ∇pf
i
∇xU · ∇pf i ∇x(U − E) · ∇pf−
,
[p · ∇xΘ Θ†, F ′] = ξ
2<f i f− − f+
f− − f+ −2<f i
,
Plug them back in (A.3.2), up to the first order in ε, one gets (3.2.18).
180
Appendix B
Kinetic theory
B.1 Positivity of the mass density in ExpRK-V
Theorem B.1.1. The method ExpRK-V defined by (5.2.8) gives positive ρ, and the negative
part of T is at most of order O(hε).
To prove this theorem, we firstly check the following lemma.
Lemma B.1.1. In each sub-stage, the distribution function f (i) and M (i) have the same first
d+ 2 moments.
Proof. We prove this for sub-stage i. Assume for ∀j < i, one has
∫
1
v
v2
2
(f (j) −M (j))dv = 0. (B.1.1)
181
Then, one could take moments of the first equation in the scheme (5.2.8a), and gets
∫
1
v
v2
2
(f (i) −M (i))eciλdv =
∫
1
v
v2
2
(fn −Mn) dv (B.1.2a)
+i−1∑j=1
aijλ
µecjλ
∫
1
v
v2
2
(P (j) − µM (j)
)dv (B.1.2b)
−i−1∑j=1
aijλ
µecjλ
∫
1
v
v2
2
(εv · ∇xf (j) − ε∂tM (j)
)dv
(B.1.2c)
(B.1.2a) is zero for sure, (B.1.2b) is zero by definition of P and (B.1.1). (B.1.2c) is zero
because of the computation from (5.2.11). Thus it is obvious that f (i) and M (i) share the same
moments on each stage.
With the previous lemma in hand, one could prove Theorem 4.
Proof. As in the previous lemma, we only do the proof for sub-stage i. The final step can be
dealt with in the same way. Rewrite the second equation of (5.2.8a) in Shu-Osher representation
∫φf (i)dv =
i−1∑j=1
(αij
∫φf (j)dv + βijh
∫φv · ∇xf (j)dv
)(B.1.3)
This moment equation is the same as the equation on ρ in the Euler system, and the classical
proof for ρ being positive for the Euler equation can just be adopted [83]. To check the
182
positivity of T , one just need to make use of the last line of the moment equation, i.e.
∫v2
2f (i)dv =
i−1∑j=1
(αij
∫v2
2f (j)dv + βijh
∫v2
2v · ∇xf (j)dv
)
=
i−1∑j=1
(αij
∫v2
2f (j)dv + βijh
∫v2
2v · ∇xM (j)dv
)(B.1.4a)
+ h
i−1∑j=1
βij
∫v2
2v · ∇x
(f (j) −M (j)
)dv (B.1.4b)
(B.1.4a) is exactly what one could get when computing for E in the Euler system: the form of
M closes it up. So the classical method to prove that E > ρu2
2 in Runge-Kutta scheme could
be used, and the only thing new is from (B.1.4b). However, as proved in the section about
AP, the difference between f and M is at most of ε, thus (B.1.4b) is of order O(hε).
B.2 ‖P (f)− P (g)‖ ≤ ‖f − g‖ for single species
B.2.1 In d2 norm
We adopt the results from [183]. They denote P2 the collection of distributions F such
that ∫Rd|v|2dF (v) <∞
A metric d2 on P2 is defined by
d2(F,G) = supξf(ξ)− g(ξ)
|ξ|2 (B.2.1)
where f is the Fourier transform of F
f(ξ) =
∫e−iξ·vdF (v)
183
One can transform the Boltzmann equation into its Fourier space and obtains[170, 20]
∂tf(t, ξ) =
∫S2
B
(ξ · n|ξ|
)[f(ξ+)f (ξ−)− f (ξ)f(0)
]dn (B.2.2)
where ξ± = ξ±|ξ|n2
Theorem B.2.1. d2(Pf , Pg) < d2(f, g) for Maxwell molecules with cut-off collision kernel.
Proof. For Maxwell molecule with cut-off collision kernel∫B = S. Thus
sup|Q−| = sup
∣∣∣∣∫ Bf∗dΩdv∗
∣∣∣∣ = sup|ρS| <∞.
Considering P = Q + µf = Q+ + (µ−Q−) f , it is enough to prove d2(Q+f , Q
+g ) < Cd2(f, g)
for C big enough. Given
Q+f =
∫S2
B
(ξ · n|ξ|
)[f(ξ+)f (ξ−)
]dn,
one has
Q+f − Q+
g
|ξ|2 =
∫S2
B
(ξ · n|ξ|
)[f(ξ+)f(ξ−)− g(ξ+)g(ξ−)
|ξ|2
]dn
From [183], one gets ∣∣∣∣∣ f(ξ+)f(ξ−)− g(ξ+)g(ξ−)
|ξ|2
∣∣∣∣∣ ≤ sup
∣∣∣∣∣ f − g|ξ|2∣∣∣∣∣
Thus, one has:
d2(Q+f , Q
+g ) = supξ
∣∣∣∣∣Q+f − Q+
g
|ξ|2
∣∣∣∣∣ ≤ S sup
∣∣∣∣∣ f − g|ξ|2∣∣∣∣∣ = Sd2(f, g)
184
B.2.2 In L2 norm
Theorem B.2.2. Given a system with cut-off collision kernel, if ‖f‖∞ < ∞, then ‖P‖2 is
controlled by ‖f‖2.
Proof. Since µ = sup‖∫Bf∗dv∗dΩ‖ < ∞, one just need to prove that Q+ is also a bounded
operator. We only consider Maxwell molecule with cut-off collision kernel.
‖Q+(f)‖2 =
∫ (∫ ∫B(Ω)f ′f ′∗dv∗dΩ
)2
dv
≤∫ ∫ ∫ (
B(Ω)f ′f ′∗)2dv∗dΩdv
≤∫B2(Ω)dΩ
∫(f ′)2dv′
∫(f ′∗)
2dv′∗
≤ C‖f‖22
≤ C√‖f‖∞‖f‖1‖f‖2
In this derivation, we used that:
1. the Hessian matrix generated by changing of variable from (v, v∗) to (v′, v∗) has deter-
minant one;
2. the collision kernel is a cut-off collision;
3. Holder Inequality.
Given ‖f‖1 = ρ, under the assumption that ‖f‖∞ is bounded, one has that Q+ is a bounded
operator on ‖f‖.
185
B.3 ‖P (f)‖ < C‖f‖ multi-species
B.3.1 In d2 norm
This is extension from section B.2.1. Here in multipspecies system, the Fourier transform
of the collision term is [19]:
∂tfi(t, ξ) =∑j
∫S2
Bij
(ξ · n|n|
)[f(ξ+
ij )f(ξ−ij )− f(ξ)f(0)]dn, (B.3.1)
where ξ+ij =
miξ+mjξ|ω|mi+mj
and ξ−ij =mj
mi+mj(ξ− |ξ|ω). Unlike in single species case, here, one only
has the momentum conservation, which is reflected as:
ξ = ξ+ij + ξ−ij (B.3.2a)
|ξ|2 6= |ξ+|2 + |ξ−|2 = |ξ|2(
1− 4mimjs
(mi +mj)2+
4m2js
(mi +mj)2
)(B.3.2b)
with s = 12
(1− ξ·ω
|ξ|
). By definition, s ∈ [0, 1], obviously, the following is bounded, when the
system is given and the mass ratio is fixed:
0 ≤|ξ−ij |2 + |ξ+
ij |2|ξ|2 ≤ Cij . (B.3.3)
Define C = maxi,jCij , then we have the theorem:
Theorem B.3.1. In multispecies case, there exists a constant C big enough such that
d2(Pf , Pg) < Cd2(f ,g)
for Maxwell molecule with cut-off collision.
Here d2 for vector functions is defined by: d2(f ,g) = maxid2(fi, gi).
186
Proof. As in the previous theorem, one only need to prove:
d2(Q+(f), Q+(g)) ≤ Cd2(f ,g). (B.3.4)
This is given by the following:
∣∣∣∣∣Q+i (f)− Q+
i (g)
|ξ|2
∣∣∣∣∣ =
∣∣∣∣∣∣∑j
∫σij
fi(ξ+ij )fj(ξ
−ij )− gi(ξ+
ij )gj(ξ−ij )
|ξ|2 dn
∣∣∣∣∣∣≤∑j
∫σij
(f+i
∣∣∣∣∣ f−j − g−j|ξ−ij |2
∣∣∣∣∣ ·∣∣∣∣∣ |ξ−ij |2|ξ|2
∣∣∣∣∣+ g−j
∣∣∣∣∣ f+i − g+
i
|ξ+ij |2
∣∣∣∣∣ ·∣∣∣∣∣ |ξ
+ij |2|ξ|2
∣∣∣∣∣)
dn
≤∑j
∫σij · supξ,j
∣∣∣∣∣ fj − gj|ξ|2
∣∣∣∣∣ · |ξ+ij |2 + |ξ−ij |2
|ξ|2dn
≤ C · S · supξ,j
∣∣∣∣∣ fj − gj|ξ|2
∣∣∣∣∣ (B.3.5)
For convenience, denote f+i = fi(ξ
+ij ) and f−j = fi(ξ
−ij ). In the derivation above, the following
are used:
• collision kernel is cut-off, thus∑
j
∫σijdn = S
• |ξ+ij |
2+|ξ−ij |2
|ξ|2 is controlled as in (B.3.3)
(B.3.5) is true for all ξ and i, thus one has:
supξ,i
∣∣∣∣∣Q+i (f)− Q+
i (g)
|ξ|2
∣∣∣∣∣ ≤ C · S · supξ,i
∣∣∣∣∣ fi − gi|ξ|2
∣∣∣∣∣ (B.3.6)
Choose C = C · S to finish the proof.
Remark B.3.1. The theorem above is based on the assumption the new metrics is well defined.
This is true since the triangular inequality still holds. The proof is straightforward and we will
skip the details here.
187
B.3.2 In L2 norm
It is an easy extension from section B.2.2. Here Qi =∑
ikQik.
Theorem B.3.2. Given a system of N species, with cut-off collision kernel, if ‖f‖∞ < ∞,
then ‖P‖2 = supi‖Pi‖2 is controlled by ‖f‖2 = supi‖fi‖2.
Proof. To prove Q+i is a bounded operator on f , we have the following:
‖Q+i (f)‖22 =
∫ (∑k
∫ ∫Bik(Ω)f ′if
′k∗dv∗dΩ
)2
dv
≤∫ ∫ ∫ ∑
k
(Bik(Ω)f ′if
′k∗)2dv∗dΩdv
≤∑k
(∫B2ik(Ω)dΩ
∫(f ′i)
2dv′∫
(f ′k∗)2dv′∗
)≤ C
∑k
‖fi‖22‖fk‖22
≤(C∑k
‖fk‖∞‖fk‖1)‖fi‖22
Thus there is a constant C big enough such that ∀i, ‖Q−i (f)‖2 < C‖fi‖2. Given Q−i is a
bounded linear functional on fi and µ <∞, one can easily get
‖Pi(f)‖2 < C‖fi‖2,
thus
supi‖Pi(f)‖2 < Csupi‖fi‖2.
188
Bibliography
[1] P. Andries, K. Aoki, and B. Perthame, A consistent BGK-type model for gas
mixtures, Journal of Statistical Physics, 106 (2002), pp. 993–1018.
[2] M. Baer and R. Englman, A study of the diabatic electronic representation within the
Born-Oppenheimer approximation, Molecular Physics, 75 (1992), pp. 293–303.
[3] G. Bal, A. Fannjiang, G. Papanicolaou, and L. Ryzhik, Radiative transport in a
periodic structure, Journal of Statistical Physics, 95 (1999), pp. 479–494.
[4] C. Bardos, R. Caflisch, and B. Nicolaenko, The Milne and Kramers problems
for the Boltzmann equation of a hard sphere gas, Communications on Pure and Applied
Mathematics, 39 (1986), pp. 323–352.
[5] C. Bardos, F. Golse, and C. Levermore, Fluid dynamic limits of the Boltzmann
equation I, Journal of Statistical Physics, 63 (1991), pp. 323–344.
[6] C. Bardos, F. Golse, and C. D. Levermore, Fluid dynamic limits of kinetic equa-
tions II: convergence proofs for the Boltzmann equation, Communications on Pure and
Applied Mathematics, 46 (1993), pp. 667–753.
[7] C. Bardos, F. Golse, and Y. Sone, Half-space problems for the Boltzmann equation:
A survey, Journal of Statistical Physics, 124 (2006), pp. 275–300.
[8] C. Bardos and S. Ukai, The classical incompressible Navier-Stokes limit of the Boltz-
mann equation, Mathematical Models and Methods in Applied Sciences, 01 (1991),
pp. 235–257.
189
[9] P. Bechouche, Semi-classical limits in a crystal with a coulombian self-consistent po-
tential: Effective mass theorems, Asymptotic Analysis, 19 (1999), pp. 95–116.
[10] P. Bechouche, N. J. Mauser, and F. Poupaud, Semiclassical limit for the
Schrodinger-Poisson equation in a crystal, Communications on Pure and Applied Math-
ematics, 54 (2001), pp. 851–890.
[11] P. Bechouche and F. Poupaud, Semi-classical limit of a Schrodinger equationfor a
stratified material, Monatshefte fur Mathematik, 129 (2000), pp. 281–301.
[12] M. V. Berry, Quantal phase factors accompanying adiabatic changes, Proceedings of the
Royal Society of London. A. Mathematical and Physical Sciences, 392 (1984), pp. 45–57.
[13] , Geometric amplitude factors in adiabatic quantum transitions, Proceedings of the
Royal Society of London. Series A: Mathematical and Physical Sciences, 430 (1990),
pp. 405–411.
[14] C. Besse, S. Borghol, T. Goudon, I. Lacroix-Violet, and J.-P. Dudon, Hy-
drodynamic regimes, Knudsen layer, numerical schemes: Definition of boundary fluxes,
Advances in Applied Mathematics and Mechanics, 3 (2011), pp. 519–561.
[15] P. L. Bhatnagar, E. P. Gross, and M. Krook, A model for collision processes
in gases. I. small amplitude processes in charged and neutral one-component systems,
Physical Review, 94 (1954), pp. 511–525.
[16] G. Billing and S. Adhikari, The time-dependent discrete variable representation
method in molecular dynamics, Chemical Physics Letters, 321 (2000), pp. 197 – 204.
[17] F. Bloch, Uber die quantenmechanik der elektronen in kristallgittern, Zeitschrift fur
Physik A Hadrons and Nuclei, 52 (1929), pp. 555–600.
190
[18] A. Bobylev, J. Carrillo, and I. Gamba, On some properties of kinetic and hydro-
dynamic equations for inelastic interactions, Journal of Statistical Physics, 98 (2000),
pp. 743–773.
[19] A. Bobylev and I. Gamba, Boltzmann equations for mixtures of Maxwell gases: Exact
solutions and power like tails, Journal of Statistical Physics, 124 (2006), pp. 497–516.
10.1007/s10955-006-9044-8.
[20] A. V. Bobylev, The Fourier transform method in the theory of the Boltzmann equation
for Maxwellian molecules, Akademiia Nauk SSSR Doklady, 225 (1975), pp. 1041–1044.
[21] A. Bohm, A. Mostafazadeh, H. Koizumi, Q. Niu, and J. Zwanziger, The Geomet-
ric Phase in Quantum Systems: Foundations, Mathematical Concepts, and Applications
in Molecular and Condensed Matter Physics (Theoretical and Mathematical Physics),
Texts and monographs in physics, Springer, 2003.
[22] M. Born and R. Oppenheimer, Zur quantentheorie der molekeln, Annalen der Physik,
389 (1927), pp. 457–484.
[23] J. Bourgat, P. Le Tallec, B. Perthame, and Y. Qiu, Coupling Boltzmann and
Euler equations without overlapping, in Domain Decomposition Methods in Science and
Engineering, A. Quarteroni, J. Periaux, Y. A. Kuznetsov, and O. Widlund, eds., Provi-
dence, RI, 1994, AMS, pp. 377–398.
[24] R. Bourquin, V. Gradinaru, and G. Hagedorn, Non-adiabatic transitions near
avoided crossings: theory and numerics, Journal of Mathematical Chemistry, 50 (2012),
pp. 602–619.
[25] S. I. Braginskii, Transport processes in a plasma, Reviews of Plasma Physics, 1 (1965),
p. 205.
191
[26] W. Braun and K. Hepp, The Vlasov dynamics and its fluctuations in the 1/n limit
of interacting classical particles, Communications in Mathematical Physics, 56 (1977),
pp. 101–113.
[27] H. Breuer and F. Petruccione, The Theory of Open Quantum Systems, Clarendon
Press, 2002.
[28] L. Butler, Chemical reaction dynamics beyond the Born-Oppenheimer approximation,
Annual Review of Physical Chemistry, 49 (1998), pp. 125–171.
[29] R. Caflisch, S. Jin, and G. Russo, Uniformly accurate schemes for hyperbolic systems
with relaxations, SIAM Journal of Numerical Analysis, 34 (1997), pp. 246–281.
[30] R. E. Caflisch, The fluid dynamic limit of the nonlinear Boltzmann equation, Com-
munications on Pure and Applied Mathematics, 33 (1980), pp. 651–666.
[31] R. Carles, P. Markowich, and C. Sparber, Semiclassical asymptotics for weakly
nonlinear Bloch waves, Journal of Statistical Physics, 117 (2004), pp. 343–375.
[32] J. A. Carrillo, C. Cercignani, and I. M. Gamba, Steady states of a Boltzmann
equation for driven granular media, Physical Review E, 62 (2000), pp. 7700–7707.
[33] C. Cercignani, The Boltzmann equation and its applications, Applied mathematical
sciences, Springer-Verlag, 1988.
[34] C. Cercignani, R. Illner, and M. Pulvirenti, The Mathematical Theory of Dilute
Gases, no. v. 106 in Applied Mathematical Sciences Series, Springer-Verlag GmbH, 1994.
[35] L. Chai, S. Jin, and Q. Li, Semi-classical limit of the Schrodinger equation with
periodic potential and band-crossing, Kinetic and Related Models, (to appear).
192
[36] L. Chai, S. Jin, Q. Li, and O. Morandi, Energy surface hopping problem in molecular
dynamics: a multiband Wigner approach, (preprint).
[37] S. Chapman and T. Cowling, The Mathematical Theory of Non-uniform Gases: An
Account of the Kinetic Theory of Viscosity, Thermal Conduction and Diffusion in Gases,
Cambridge Mathematical Library, Cambridge University Press, 1970.
[38] S. Child, Molecular collision theory, Dover Books on Chemistry Series, Dover Publica-
tions, Incorporated, 1974.
[39] R. M. Chmieleski and J. H. Ferziger, Transport properties of a nonequilibrium
partially ionized gas, Physics of Fluids, 10 (1967), pp. 364–371.
[40] L. Cohen, Generalized phase-space distribution functions, Journal of Mathematical
Physics, 7 (1966), pp. 781–786.
[41] J.-M. Combes, P. Duclos, and R. Seller, The Born-Oppenheimer approximation,
in Rigorous atomic and molecular physics, G. Velo and A. S. Wightman, eds., Plenum
Press, 1981, pp. 185–212.
[42] F. Coron and B. Perthame, Numerical passage from kinetic to fluid equations, SIAM
Journal of Numerical Analysis, 28 (1991), pp. 26–42.
[43] R. Courant and D. Hilbert, Methods of Mathematical Physics: Volume 2, Wiley
India Pvt. Ltd., 2008.
[44] A. de Masi, R. Esposito, and J. L. Lebowitz, Incompressible Navier-Stokes and
Euler limits of the Boltzmann equation, Communications on Pure and Applied Mathe-
matics, 42 (1989), pp. 1189–1214.
[45] P. Degond and S. Jin, A smooth transition model between kinetic and diffusion equa-
tions, SIAM Journal Numerical Analysis, 42 (2005), pp. 2671–2687.
193
[46] P. Degond, S. Jin, and L. Mieussens, A smooth transition model between kinetic
and hydrodynamic equations, Journal Computional Physics, 209 (2005), pp. 665–694.
[47] P. Degond and B. Lucquin-Desreux, The Fokker-Planck asymptotics of the Boltz-
mann collision operator in the Culomb case, Mathematical Models and Methods in Ap-
plied Sciences, 02 (1992), pp. 167–182.
[48] P. Degond and B. Lucquin-Desreux, The asymptotics of collision operators for two
species of particles of disparate masses, Mathematical Models and Methods in Applied
Sciences, 6 (1996), pp. 405–436.
[49] , Transport coefficients of plasmas and disparate mass binary gases, Transport The-
ory and Statistical Physics, 25 (1996), pp. 595–633.
[50] L. Desvillettes, On asymptotics of the Boltzmann equation when the collisions become
grazing, Transport Theory and Statistical Physics, 21 (1992), pp. 259–276.
[51] A. S. Dickinson and P. R. Certain, Calculation of matrix elements for one-
dimensional quantum-mechanical problems, The Journal of Chemical Physics, 49 (1968),
pp. 4209–4211.
[52] G. Dimarco and L. Pareschi, Exponential Runge–Kutta methods for stiff kinetic
equations, SIAM Journal Numerical Analysis, 49 (2011), pp. 2057–2077.
[53] , High order asymptotic-preserving schemes for the Boltzmann equation, Comptes
Rendus Mathematique, 350 (2012), pp. 481–486.
[54] R. J. DiPerna and P.-L. Lions, On the Cauchy problem for Boltzmann equations:
Global existence and weak stability, Annals of Mathematics, 130 (1989), pp. pp. 321–366.
[55] K. Drukker, Basics of surface hopping in mixed quantum/classical simulations, Journal
of Computational Physics, 153 (1999), pp. 225–272.
194
[56] J. Duistermaat, Fourier integral operators, Progress in Mathematics Series, Birkhauser
Verlag GmbH, 1996.
[57] W. E, J. Lu, and X. Yang, Asymptotic analysis of quantum dynamics in crystals: The
Bloch-Wigner transform, Bloch dynamics and Berry phase, Acta Mathematicae Appli-
catae Sinica, English Series, (2011), pp. 1–12.
[58] C. Eckart, The kinetic energy of polyatomic molecules, Physical Review, 46 (1934),
pp. 383–387.
[59] R. Ellis and M. Pinsky, The first and second fluid approximations to the linearized
Boltzmann equation, Journal de Mathematiques Pures et Appliquees, 54 (1975), pp. 125–
156.
[60] C. Fermanian Kammerer, Wigner measures and molecular propagation through
generic energy level crossings, Reviews in Mathematical Physics, 15 (2003), pp. 1285–
1317.
[61] C. Fermanian Kammerer and P. Gerard, A Landau-Zener formula for non-
degenerated involutive codimension 3 crossings, Annales Henri Poincare, 4 (2003),
pp. 513–552.
[62] C. Fermanian Kammerer and C. Lasser, Wigner measures and codimension two
crossings, Journal of Mathematical Physics, 44 (2003), pp. 507–527.
[63] J. H. Ferziger and H. G. Kaper, Mathematical Theory of Transport Processes in
Gases, North-Holland Pub. Co, 1972.
[64] F. Filbet, J. Hu, and S. Jin, A numerical scheme for the quantum Boltzmann equation
efficient in the fluid regime, Mathematical Modelling and Numerical Analysis, 46 (2012),
pp. 443–463.
195
[65] F. Filbet and S. Jin, A class of asymptotic-preserving schemes for kinetic equa-
tions and related problems with stiff sources, Journal Computional Physics, 229 (2010),
pp. 7625–7648.
[66] , An asymptotic preserving scheme for the ES-BGK model of the Boltzmann equa-
tion, SIAM Journal Scientifical Computing, 46 (2011), pp. 204–224.
[67] F. Filbet and C. Mouhot, Analysis of spectral methods for the homogeneous
Boltzmann equation, Transactions of the American Mathematical Society, 363 (2011),
pp. 1947–1980.
[68] F. Filbet and A. Rambaud, Analysis of an asymptotic preserving scheme for re-
laxation systems, ESAIM: Mathematical Modelling and Numerical Analysis, 47 (2013),
pp. 609–633.
[69] G. Floquet, Sur les equations differentielles lineaires a coefficients periodiques, Annales
scientifiques de l’Ecole Normale Superieure, 12 (1883), pp. 47–88.
[70] E. Gabetta, L. Pareschi, and G. Toscani, Relaxation schemes for nonlinear kinetic
equations, SIAM Journal of Numerical Analysis, 34 (1997), pp. 2168–2194.
[71] I. Gamba, V. Panferov, and C. Villani, On the Boltzmann equation for diffusively
excited granular media, Communications in Mathematical Physics, 246 (2004), pp. 503–
541.
[72] V. Garzo, A. Santos, and J. J. Brey, A kinetic model for a multicomponent gas,
Physics of Fluids, 1 (1989), pp. 380–384.
[73] P. Gerard, Mesures semi-classiques et ondes de Bloch, Seminaire Equations aux
derivees partielles, (1991), pp. 1–19.
196
[74] P. Gerard, P. Markowich, N. Mauser, and F. Poupaud, Homogenization limits
and Wigner transforms, Communications on Pure and Applied Mathematics, 50 (1997),
pp. 323–379.
[75] E. Goldman and L. Sirovich, Equations for gas mixtures, Physics of Fluids, 10 (1967),
pp. 1928–1940.
[76] F. Golse, S. Jin, and C. D. Levermore, A domain decomposition analysis for a
two-scale linear transport problem, Mathematical Modeling and Numerical Analysis, 37
(2002), pp. 869–892.
[77] F. Golse and L. Saint-Raymond, The Navier-Stokes limit of the Boltzmann equation
for bounded collision kernels, Inventiones mathematicae, 155 (2004), pp. 81–161.
[78] , The Navier-Stokes limit of the Boltzmann equationfor hard potentials, Journal de
Mathematiques Pures et Appliquees, 91 (2009), pp. 508–552.
[79] L. Gosse, Multiphase semiclassical approximation of an electron in a one-dimensional
crystalline lattice II. impurities, confinement and Bloch oscillations, Journal of Compu-
tational Physics, 201 (2004), pp. 344 – 375.
[80] , A case study on the reliability of multiphase WKB approximation for the one-
dimensional Schrodinger equation, in Numerical Methods for Hyperbolic and Kinetic
Problems, S. Cordier, T. Goudon, M. Gutnic, and E. Sonnendrucker, eds., European
Mathematical Society Publishing House, 2005.
[81] L. Gosse and P. A. Markowich, Multiphase semiclassical approximation of an elec-
tron in a one-dimensional crystalline lattice: I. homogeneous problems, Journal of Com-
putational Physics, 197 (2004), pp. 387 – 417.
197
[82] L. Gosse and N. Mauser, Multiphase semiclassical approximation of an electron in
a one-dimensional crystalline lattice-III. from ab initio models to wkb for Schrodinger-
Poisson, Journal of Computational Physics, 211 (2006), pp. 326–346.
[83] S. Gottlieb and C. wang Shu, Total variation diminishing Runge-Kutta schemes,
Mathematics of Computation, 67 (1998), pp. 73–85.
[84] S. Gottlieb, C. wang Shu, and E. Tadmor, Strong stability-preserving high-order
time discretization methods, SIAM Review, 43 (2001), pp. 89–112.
[85] H. Grad, Principles of the Kinetic Theory of Gases, Handbuch der Physik, 12 (1958),
pp. 205–294.
[86] H. Grad, Theory of rarefied gases, in Rarefied Gas Dynamics, F. Devienne, ed., Perga-
mon, London, England, 1960, pp. 100–138.
[87] H. Grad, Asymptotic Theory of the Boltzmann Equation, II, in Rarefied Gas Dynamics,
Volume 1, J. A. Laurmann, ed., 1963, p. 26.
[88] E. P. Gross and M. Krook, Model for collision processes in gases: Small-amplitude
oscillations of charged two-component systems, Physical Review, 102 (1956), pp. 593–604.
[89] J. Guillot, J. Ralston, and E. Trubowitz, Semi-classical asymptotics in solid
state physics, Communications in Mathematical Physics, 116 (1988), pp. 401–415.
[90] G. Hagedorn, A time dependent Born-Oppenheimer approximation, Communications
in Mathematical Physics, 77 (1980), pp. 1–19.
[91] , Proof of the Landau-Zener formula in an adiabatic limit with small eigenvalue
gaps, Communications in Mathematical Physics, 136 (1991), pp. 433–449.
198
[92] G. Hagedorn, Classification and normal forms for avoided crossings of quantum-
mechanical energy levels, Journal of Physics A: Mathematical and General, 31 (1998),
p. 369.
[93] G. A. Hagedorn, High order corrections to the time-dependent Born-Oppenheimer
approximation I: Smooth potentials, Annals of Mathematics, 124 (1986), pp. pp. 571–
590.
[94] G. A. Hagedorn, High order corrections to the time-dependent Born-Oppenheimer
approximation. II: Coulomb systems, Communications in Mathematical Physics, 117
(1988), pp. 387–403.
[95] G. A. Hagedorn and A. Joye, Landau-Zener transitions through small electronic
eigenvalue gaps in the Born-Oppenheimer approximation, Annales de l’institut Henri
Poincare (A) Physique theorique, 68 (1998), pp. 85–134.
[96] , Molecular propagation through small avoided crossings of electron energy levels,
Reviews in Mathematical Physics, 11 (1999), pp. 41–101.
[97] , A time-dependent BornOppenheimer approximation with exponentially small error
estimates, Communications in Mathematical Physics, 223 (2001), pp. 583–626.
[98] E. Hairer, S. Nørsett, and G. Wanner, Solving Ordinary Differential Equations I:
Nonstiff Problems (Springer Series in Computational Mathematics), Springer, 2010.
[99] D. O. Harris, G. G. Engerholm, and W. D. Gwinn, Calculation of matrix elements
for one-dimensional quantum-mechanical problems and the application to anharmonic
oscillators, Journal of Chemical Physics, 43 (1965), pp. 1515–1517.
[100] H. Haug and A. Jauho, Quantum Kinetics in Transport and Optics of Semiconductors,
Springer Series in Solid-state Sciences, v. 123, Springer London, Limited, 2008.
199
[101] D. Hilbert, Begrundung der kinetischen gastheorie, Mathematische Annalen, 72 (1912),
pp. 562–577.
[102] N. Hill, Gaussian beam migration, Geophysics, 55 (1990), pp. 1416–1428.
[103] L. Hormander, The Weyl calculus of pseudo-differential operators, Communications
on Pure and Applied Mathematics, 32 (1979), pp. 359–443.
[104] J. Hu, S. Jin, and B. Yan, A numerical scheme for the quantum Fokker-Planck-Landau
equation efficient in the fluid regime, Communications in Computational Physics, (to
appear).
[105] Z. Huang, S. Jin, P. Markowich, and C. Sparber, A Bloch decomposition-based
split-step pseudospectral method for quantum dynamics with periodic potentials, SIAM
Journal on Scientific Computing, 29 (2008), pp. 515–538.
[106] Z. Huang, S. Jin, P. Markowich, and C. Sparber, Numerical simulation of the non-
linear Schrodinger equation with multidimensional periodic potentials, Multiscale Mod-
eling & Simulation, 7 (2008), pp. 539–564.
[107] Z. Huang, S. Jin, H. Wu, and D. Yin, Gaussian beam methods for the Dirac equa-
tion in the semi-classical regime, Communications in Mathematical Sciences, 10 (2012),
pp. 1301–1315.
[108] V. Jaksic and J. Segert, Exponential approach to the adiabatic limit and the Landau-
Zener formula, Reviews in Mathematical Physics, 04 (1992), pp. 529–574.
[109] S. Jin, Runge-Kutta methods for hyperbolic conservation laws with stiff relaxation terms,
Journal of Computational Physics, 122 (1995), pp. 51–67.
[110] , Efficient asymptotic-preserving (AP) schemes for some multiscale kinetic equa-
tions, SIAM Journal of Scientific Computing, 21 (1999), pp. 441–454.
200
[111] , Asymptotic preserving (AP) schemes for multiscale kinetic and hyperbolic equa-
tions: a review, in Lecture Notes for Summer School on ”Methods and Models of Kinetic
Theory” (M&MKT), R. di Matematica della Universita di Parma, ed., Porto Ercole
(Grosseto, Italy), June 2010.
[112] S. Jin and Q. Li, A BGK-penalization-based asymptotic-preserving scheme for the
multi-species Boltzmann equation, Numerical Methods for Partial Differential Equations,
29 (2013), pp. 1056–1080.
[113] S. Jin and X. Li, Multi-phase computations of the semiclassical limit of the Schrodinger
equation and related problems: Whitham vs Wigner, Physica D: Nonlinear Phenomena,
182 (2003), pp. 46 – 85.
[114] S. Jin, H. Liu, S. Osher, and Y. Tsai, Computing multivalued physical observables for
the semiclassical limit of the Schrodinger equation, Journal of Computational Physics,
205 (2005), pp. 222–241.
[115] S. Jin, P. Markowich, and C. Sparber, Mathematical and computational methods
for semiclassical Schrodinger equations, Acta Numerica, 20 (2011), pp. 121–209.
[116] S. Jin and S. Osher, A level set method for the computation of multivalued solutions to
quasi-linear hyperbolic PDEs and Hamilton-Jacobi equations, Communications in Math-
ematical Sciences, 1 (2003), pp. 575–591.
[117] S. Jin and P. Qi, A hybrid Schrodinger/Gaussian beam solver for quantum barriers
and surface hopping, Kinetic and Related Models, 4 (2011), pp. 1097–1120.
[118] S. Jin, P. Qi, and Z. Zhang, An Eulerian surface hopping method for the Schrodinger
equation with conical crossings, SIAM Multiscale Modeling and Simulation, 9 (2011),
pp. 258–281.
201
[119] S. Jin and Y. Shi, A micro-macro decomposition-based asymptotic-preserving scheme
for the multispecies Boltzmann equation, SIAM Journal of Scientific Computing, 31
(2010), pp. 4580–4606.
[120] S. Jin, H. Wu, and X. Yang, Gaussian beam methods for the Schrodinger equation
in the semi-classical regime: Lagrangian and Eulerian formulations, Communications in
Mathematical Sciences, 6 (2008), pp. 995–1020.
[121] , A numerical study of the Gaussian beam methods for Schrodinger-Poisson equa-
tions, Journal of Computational Mathematics, 28 (2010), p. 261.
[122] , Semi-Eulerian and high order Gaussian beam methods for the Schrodinger equa-
tion in the semiclassical regime, Communications in Computational Physics, 9 (2011),
pp. 668–687.
[123] S. Jin and B. Yan, A class of asymptotic-preserving schemes for the Fokker-Planck-
Landau equation, Journal of Computational Physics, 230 (2011), pp. 6420–6437.
[124] S. Jin and X. Yang, Computation of the semiclassical limit of the Schrodinger equation
with phase shift by a level set method, Journal of Scientific Computing, 35 (2008), pp. 144–
169.
[125] S. Jin and D. Yin, Computational high frequency waves through curved interfaces via
the Liouville equation and geometric theory of diffraction, Journal of Computational
Physics, 227 (2008), pp. 6106–6139.
[126] E. A. Johnson, Energy and momentum equations for disparate-mass binary gases,
Physics of Fluids, 16 (1973), pp. 45–49.
[127] A. Joye, Proof of the Landau-Zener formula, Asymptotic Analysis, 9 (1994), pp. 209–
258.
202
[128] A. Klar, H. Neunzert, and J. Struckmeier, Transition from kinetic theory to
macroscopic fluid equations: A problem for domain decomposition and a source for new
algorithms, Transport Theory and Statistical Physics, 29 (2000), pp. 93–106.
[129] I. Kodi Husim, Some Formal Properties of the Density Matrix, Nippon Sugaku-
Buturigakkwai Kizi Dai 3 Ki, 22 (1940), pp. 264–314.
[130] S. Kube, C. Lasser, and M. Weber, Monte Carlo sampling of Wigner functions
and surface hopping quantum dynamics, Journal of Computational Physics, 228 (2009),
pp. 1947 – 1962.
[131] L. Landau, Zur Theorie der Energieubertragung. II., Physics of the Soviet Union2,
(1932), pp. 46–51.
[132] O. E. Lanford, Time evolution of large classical systems, in Dynamical Systems, The-
ory and Applications, J. Moser, ed., vol. 38 of Lecture Notes in Physics, Springer Berlin
Heidelberg, 1975, pp. 1–111.
[133] , The hard sphere gas in the Boltzmann-grad limit, Physica A: Statistical Mechanics
and its Applications, 106 (1981), pp. 70 – 76.
[134] E. W. Larsen, J. Morel, and W. F. Miller, Asymptotic solutions of numerical
transport problems in optically thick, diffusive regimes, Journal of Computational Physics,
69 (1987), pp. 283 – 324.
[135] E. W. Larsen and J. E. Morel, Asymptotic solutions of numerical transport problems
in optically thick, diffusive regimes II, Journal of Computational Physics, 83 (1989),
pp. 212–236.
203
[136] C. Lasser, T. Swart, and S. Teufel, Construction and validation of a rigorous sur-
face hopping algorithm for conical crossings, Communications in Mathematical Sciences,
5 (2007), pp. 789–814.
[137] C. Lasser and S. Teufel, Propagation through conical crossings: an asymptotic semi-
group, Communications on Pure and Applied Mathematics, 58 (2005), pp. 1188–1230.
[138] P. Le Tallec and F. Mallinger, Coupling Boltzmann and Navier-Stokes equations
by half fluxes, Journal of Computational Physics, 136 (1997), pp. 51–67.
[139] M. Lemou and L. Mieussens, A new asymptotic preserving scheme based on micro-
macro formulation for linear kinetic equations in the diffusion limit, SIAM Journal of
Scientific Computing, 31 (2008), pp. 334–368.
[140] J. Leray, etude de diverses equations integrales non lineaires et quelques problemes
que pose l’hydrodynamique, Journal de Mathematiques Pures et Appliquees, 9 (1933),
pp. 1–82.
[141] R. J. LeVeque, Numerical Methods for Conservation Laws, Lectures in mathematics
ETH Zurich, Birkhauser Verlag, 1992.
[142] R. J. LeVeque and etc., Version 4.3. ”http://www.amath.washington.edu/ claw/”.
[143] Q. Li and L. Pareschi, Exponential Runge-Kutta schemes for inhomogeneous Boltz-
mann equations with high order of accuracy, arXiv:1208.2622, (preprint).
[144] Q. Li and X. Yang, Exponential Runge-Kutta methods for the multispecies Boltzmann
equation, (preprint).
[145] J. C. Light and T. Carrington, Discrete-variable representations and their utiliza-
tion, in Advances in Chemical Physics, John Wiley & Sons, Inc., 2007, pp. 263–310.
204
[146] P.-L. Lions, Mathematical Topics in Fluid Mechanics, Vol. 1: Incompressible Models,
The Clarendon Press, Oxford University Press, New York, 1996.
[147] P.-L. Lions and T. Paul, Sur les measures de Wigner, Revista Matematica Iberoamer-
icana, 9 (1993), pp. 553–618.
[148] F. London, Uber den mechanismus der homoopolaren bindung, in Probleme der mod-
ernen physik, P. Debye, ed., S. Herzel, 1928.
[149] J. Lu and X. Yang, Frozen Gaussian approximation for high frequency wave propaga-
tion, Communications in Mathematical Sciences, 9 (2011), pp. 663–683.
[150] J. Lu and X. Yang, Convergence of frozen Gaussian approximation for high-frequency
wave propagation, Communications on Pure and Applied Mathematics, 65 (2012),
pp. 759–789.
[151] P. Markowich, N. Mauser, and F. Poupaud, A Wigner-function approach to (semi)
classical limits: Electrons in a periodic potential, Journal of Mathematical Physics, 35
(1994), pp. 1066–1094.
[152] C. A. Mead and D. G. Truhlar, Conditions for the definition of a strictly dia-
batic electronic basis for molecular systems, The Journal of Chemical Physics, 77 (1982),
pp. 6090–6098.
[153] O. Morandi, Multiband Wigner-function formalism applied to the Zener band transition
in a semiconductor, Physical Review B, 80 (2009), p. 024301.
[154] O. Morandi and F. Schurrer, Wigner model for quantum transport in graphene,
Arxiv preprint arXiv:1102.2416, (2011).
[155] H. More, R. Oppenheim, and J. Ross, The Wigner function and transport theory,
205
in Studies in Statistical Mechanics, V. D. Boer and O. E. Uhlenbeck, eds., vol. 1, North-
Holland Publishing Company, 1962.
[156] C. Mouhot and L. Pareschi, Fast algorithms for computing the Boltzmann collision
operator, Mathematics of Computation, (2004).
[157] H. Nakamura, Nonadiabatic Transition: Concepts, Basic Theories and Applications,
World Scientific Publ., 2002.
[158] H. Nakamura, Nonadiabatic chemical dynamics: Comprehension and control of dynam-
ics, and manifestation of molecular functions, in Advances in Chemical Physics, John
Wiley & Sons, Inc., 2008, pp. 95–212.
[159] T. Nishida, Fluid dynamical limit of the nonlinear Boltzmann equation to the level of
the compressible Euler equation, Communications in Mathematical Physics, 61 (1978),
pp. 119–148.
[160] F. Odeh and J. Keller, Partial differential equations with periodic coefficients and
Bloch waves in crystals, Journal of Mathematical Physics, 5 (1964), p. 1499.
[161] S. Osher and J. A. Sethian, Fronts propagating with curvature-dependent speed: al-
gorithms based on Hamilton-Jacobi formulations, Journal of Computational Physics, 79
(1988), pp. 12–49.
[162] G. Panati, H. Spohn, and S. Teufel, Effective dynamics for Bloch electrons: Peierls
substitution and beyond, Communications in Mathematical Physics, 242 (2003), pp. 547–
578.
[163] , Motion of electrons in adiabatically perturbed periodic structures, Analysis, Mod-
eling and Simulation of Multiscale Problems, (2006), pp. 595–617.
206
[164] G. Papanicolau, A. Bensoussan, and J. Lions, Asymptotic Analysis for Periodic
Structures, Studies in Mathematics and its Applications, Elsevier Science, 1978.
[165] L. Pareschi and G. Russo, Numerical solution of the Boltzmann equation I: Spectrally
accurate approximation of the collision operator, SIAM Journal of Numerical Analysis,
37 (2000), pp. 1217–1245.
[166] B. Perthame, Boltzmann type schemes for gas dynamics and the entropy property,
SIAM Journal on Numerical Analysis, 27 (1990), pp. pp. 1405–1421.
[167] S. Pieraccini and G. Puppo, Implicit-Explicit schemes for BGK kinetic equations,
Journal of Scientific Computing, 32 (2007), pp. 1–28.
[168] M. Popov, A new method of computation of wave fields using Gaussian beams, Wave
Motion, 4 (1982), pp. 85 – 97.
[169] F. Poupaud and C. Ringhofer, Semi-classical limits in a crystal with exterior poten-
tials and effective mass theorems., Communications in Partial Differential Equations, 21
(1996), pp. 1897–1918.
[170] A. Pulvirenti and G. Toscani, The theory of the nonlinear Boltzmann equation for
Maxwell molecules in Fourier representation, Annali di Matematica Pura ed Applicata,
171 (1996), pp. 181–204.
[171] J. Ralston, Gaussian beams and the propagation of singularities, in Studies in Par-
tial Differential Equations, vol. 23, Washington, DC, 1982, Mathematics Association of
America, pp. 206–248.
[172] V. Rousse, Landau-Zener transitions for eigenvalue avoided crossings in the adiabatic
and Born-Oppenheimer approsimations, Asymptotic Analysis, 37 (2004), pp. 293–328.
207
[173] L. Saint-Raymond, Hydrodynamic Limits of the Boltzmann Equation, no. no. 1971 in
Lecture Notes in Mathematics, Springer, 2009.
[174] F. Schwabl, Statistical Mechanics, Advanced Texts in Physics, Springer-Verlag GmbH,
2002.
[175] M. Sillanpaa, T. Lehtinen, A. Paila, Y. Makhlin, and P. Hakonen, Landau–
Zener interferometry in a Cooper-pair box, Journal of Low Temperature Physics, 146
(2007), pp. 253–262.
[176] R. J. Spiteri and S. J. Ruuth, A new class of optimal high-order strong-stability-
preserving time discretization methods, SIAM Journal on Numerical Analysis, 40 (2003),
pp. pp. 469–491.
[177] L. Spitzer and R. Harm, Transport phenomena in a completely ionized gas, Physical
Review, 89 (1953), pp. 977–981.
[178] H. Spohn and S. Teufel, Adiabatic decoupling and time-dependent Born-Oppenheimer
theory, Communications in Mathematical Physics, 224 (2001), pp. 113–132.
[179] P. Stanley and L. Kricka, Bioluminescence & Chemiluminescence:, World Scientific,
2002.
[180] G. Sundaram and Q. Niu, Wave-packet dynamics in slowly perturbed crystals: Gra-
dient corrections and Berry-phase effects, Physical Review B, (1999), pp. 14915–14925.
[181] N. M. Tanushev, Superpositions and higher order Gaussian beams, Communications in
Mathematical Sciences, 6 (2008), p. 449475.
[182] N. M. Tanushev, B. Engquist, and R. Tsai, Gaussian beam decomposition of high
frequency wave fields, Journal of Computational Physics, 228 (2009), pp. 8856–8871.
208
[183] G. Toscani and C. Villani, Probability metrics and uniqueness of the solution to
the Boltzmann equation for a Maxwell gas, Journal of Statistical Physics, 94 (1999),
pp. 619–637.
[184] J. Tully, Molecular dynamics with electronic transitions, The Journal of Chemical
Physics, 93 (1990), pp. 1061–1071.
[185] J. Tully and R. Preston, Trajectory surface hopping approach to nonadiabatic molec-
ular collisions: the reaction of H+ with D2, The Journal of Chemical Physics, 55 (1971),
pp. 562–572.
[186] C. Villani, Chapter 2 a review of mathematical topics in collisional kinetic theory, in
Handbook of Mathematical Fluid Dynamics, S. Friedlander and D. Serre, eds., vol. 1,
North-Holland, 2002, pp. 71 – 74.
[187] C. Villani, Mathematics of granular materials, Journal of Statistical Physics, 124
(2006), pp. 781–822.
[188] J. von Neumann and E. Wigner, Uber das verhalten von eigenwerten bei adiabatis-
chen prozessen, Zeitschrift fur Physik A Hadrons and Nuclei, 30 (1929), pp. 467–470.
[189] A. I. Voronin, J. M. C. Marques, and A. J. C. Varandas, Trajectory surface
hopping study of the Li + Li2(X1Σ g+) dissociation reaction, The Journal of Physical
Chemistry A, 102 (1998), pp. 6057–6062.
[190] C. wang Shu, Essentially non-oscillatory and weighted essentially non-oscillatory
schemes for hyperbolic conservation laws, tech. rep., Institute for Computer Applica-
tions in Science and Engineering (ICASE), 1997.
[191] C. wang Shu and S. Osher, Efficient implementation of essentially non-oscillatory
shock-capturing schemes, II, Journal of Computational Physics, 83 (1989), pp. 32–78.
209
[192] H. Weyl, Quantenmechanik und gruppentheorie, Zeitschrift fur Physik A Hadrons and
Nuclei, 46 (1927), pp. 1–46.
[193] E. Wigner, On the quantum correction for thermodynamic equilibrium, Physical Re-
view, 40 (1932), pp. 749–759.
[194] C. Wilcox, Theory of Bloch waves, Journal d’Analyse Mathematique, 33 (1978),
pp. 146–167.
[195] T. Wilson and J. W. Hastings, Bioluminescence, Annual Review of Cell and Devel-
opmental Biology, 14 (1998), pp. 197–230. PMID: 9891783.
[196] G. Worth and L. Cederbaum, Beyond Born-Oppenheimer: molecular dynamics
through a conical intersection, Annual Review of Physical Chemistry, 55 (2004), pp. 127–
158.
[197] D. Xiao, M. Chang, and Q. Niu, Berry phase effects on electronic properties, Reviews
of Modern Physics, 82 (2010), p. 1959.
[198] D. Xiao, J. Shi, and Q. Niu, Berry phase correction to electron density of states in
solids, Physical Review Letter, 95 (2005), p. 137204.
[199] B. Yan and S. Jin, A successive penalty-based asymptotic-preserving scheme for kinetic
equations, SIAM Journal on Scientific Computing, 35 (2013), pp. A150–A172.
[200] D. Yin, M. Tang, and S. Jin, The Gaussian beam method for the Wigner equation
with discontinuous potentials, Inverse Problems and Imaging, (to appear).
[201] C. Zachos, D. Fairlie, and T. Curtright, Quantum Mechanics in Phase Space: An
Overview With Selected Papers, World Scientific Series In 20th Century Physics, World
Scientific, 2005.
210
[202] C. Zener, Non-adiabatic crossing of energy levels, Proceedings of the Royal Society of
London. Series A, Containing Papers of a Mathematical and Physical Character, 137
(1932), pp. 696–702.
[203] A. Zewail, Femtochemistry: Ultrafast Dynamics of the Chemical Bond, Volume 1,
World Scientific Series in 20th Century Chemistry, World Scientific, 1994.